Environmental tobacco smoke exposure and lung cancer: A systematic review

Abstract

AIM: To review evidence relating passive smoking to lung cancer risk in never smokers, considering various major sources of bias.

METHODS: Epidemiological prospective or case-control studies were identified which provide estimates of relative risk (RR) and 95%CI for never smokers for one or more of seven different indices of exposure to environmental tobacco smoke (ETS): The spouse; household; workplace; childhood; travel; social and other; and total. A wide range of study details were entered into a database, and the RRs for each study, including descriptions of the comparisons made, were entered into a linked database. RRs were derived where necessary. Results were entered, where available, for all lung cancer, and for squamous cell cancer and adenocarcinoma. “Most adjusted” results were entered based on results available, adjusted for the greatest number of potential confounding variables. “Least adjusted” results were also entered, with a preference for results adjusted at least for age for prospective studies. A pre-planned series of fixed-effects and random-effects meta-analyses were conducted. Overall analyses and analyses by continent were run for each exposure index, with results for spousal smoking given by sex, and results for childhood exposure given by source of ETS exposure. For spousal exposure, more extensive analyses provide results by various aspects of study design and definition of the RR. For smoking by the husband (or nearest equivalent), additional analyses were carried out both for overall risk, and for risk per 10 cigarettes per day smoked by the husband. These adjusted for uncontrolled confounding by four factors (fruit, vegetable and dietary fat consumption, and education), and corrected for misclassification of smoking status of the wife. For the confounding adjustment, estimates for never smoking women were derived from publications on the relationship of the four factors to both lung cancer risk and at home ETS exposure, and on the correlations between the factors. The bias due to misclassification was calculated on the basis that the proportion of ever smokers denying smoking is 10% in Asian studies and 2.5% elsewhere, and that those who deny smoking have the same risk as those who admit it. This approach, justified in previous work, balances higher true denial rates and lower risk in deniers compared to non-deniers.

RESULTS: One hundred and two studies were identified for inclusion, published in 1981 onwards, 45 in Asia, 31 in North America, 21 in Europe, and five elsewhere. Eighty-five were of case-control design and 17 were prospective. Significant (P < 0.05) associations were noted, with random-effects of (RR = 1.22, 95%CI: 1.14-1.31, n = 93) for smoking by the husband (RR = 1.14, 95%CI: 1.01-1.29, n = 45) for smoking by the wife (RR = 1.22, 95%CI: 1.15-1.30, n = 47) for workplace exposure (RR = 1.15, 95%CI: 1.02-1.29, n = 41) for childhood exposure, and (RR = 1.31, 95%CI: 1.19-1.45, n = 48) for total exposure. No significant association was seen for ETS exposure in travel (RR = 1.34, 95%CI: 0.94-1.93, n = 8) or in social situations (RR = 1.01, 95%CI: 0.82-1.24, n = 15). A significant negative association (RR = 0.78, 95%CI: 0.64-0.94, n = 8) was seen for ETS exposure in childhood, specifically from the parents. Significant associations were also seen for spousal smoking for both squamous cell carcinoma (RR = 1.44, 95%CI: 1.15-1.80, n = 24) and adenocarcinoma (RR = 1.33, 95%CI: 1.17-1.51, n = 30). Results generally showed marked heterogeneity between studies. For smoking by either the husband or wife, where 119 RR estimates gave an overall estimate of (RR = 1.21, 95%CI: 1.14-1.29), the heterogeneity was highly significant (P < 0.001), with evidence that the largest RRs were seen in studies published in 1981-89, in small studies (1-49 cases), and for estimates unadjusted by age. For smoking by the husband, the additional analyses showed that adjustment for the four factors reduced the overall (RR = 1.22, 95%CI: 1.14-1.31) based on 93 estimates to (RR = 1.14, 95%CI: 1.06-1.22), implying bias due to uncontrolled confounding of 7%. Further correction for misclassification reduced the estimate to a marginally non-significant (RR = 1.08, 95%CI: 0.999-1.16). In the fully adjusted and corrected analyses, there was evidence of an increase in Asia (RR = 1.18, 95%CI: 1.07-1.30, n = 44), but not in other regions (RR = 0.96, 95%CI: 0.86-1.07, n = 49). Studies published in the 1980’s, studies providing dose-response data, and studies only providing results unadjusted for age showed elevated RRs, but later published studies, studies not providing dose-response data, and studies adjusting for age did not. The pattern of results for RRs per 10 cigs/d was similar, with no significant association in the adjusted and corrected results (RR = 1.03, 95%CI: 0.994-1.07).

CONCLUSION: Most, if not all, of the ETS/lung cancer association can be explained by confounding adjustment and misclassification correction. Any causal relationship is not convincingly demonstrated.

Key Words: Passive smoking, Lung neoplasms, Dose-response, Meta-analysis, Review, Confounding factors (epidemiology), Misclassification

Core tip: We present an up-to-date meta-analysis of the evidence relating non-smoker lung cancer to environmental tobacco smoke (ETS) exposure. We demonstrate a clear risk increase for spousal, at-home, workplace and total exposure, but not childhood exposure. For husband smoking, the relative risk (RR) is estimated as (RR = 1.22, 95%CI: 1.14-1.31). However, adjustment for confounding by education and dietary variables, and correction for misclassified wife’s smoking reduces it to (RR = 1.08, 95%CI: 0.999-1.16). Given the other data limitations and biases we discuss, one cannot reliably conclude that any true ETS effect on lung cancer risk exists. Our results suggest caution in drawing inferences from weak epidemiological associations where known biases exist.

INTRODUCTION

It has been widely accepted that environmental tobacco smoke (ETS) exposure increases lung cancer risk, based on various authoritative reviews[1]. However, it was suggested some years ago[2] that a substantial part, if not all, of the relationship may be due to bias resulting from confounding by other lung cancer risk factors, and misclassification of smoking habits, with some self-reported never smokers actually being smokers. While there have been various meta-analyses of the evidence in the last 20 years[3-10]. these are often limited to specific indices of exposure or regions, and typically do not include formal adjustments for potential biases. They also do not take into account all the more recent studies, with over 100 studies published by now, many relatively recent.

The objective of this review, therefore, is to present an up-to-date comprehensive meta-analysis of the available evidence which relates ETS exposure to lung cancer risk among never smokers, considering exposure from various sources, and illustrating the potential magnitude of the bias that can arise from confounding and misclassification of smoking.

MATERIALS AND METHODS

Introduction

The analyses presented were conducted in three stages. First, results of meta-analysis are presented relating a range of indices of ETS exposure to risk of lung cancer. Second, for two indices (spousal smoking and amount smoked by the spouse), individual study estimates for females are adjusted for the effects of confounding for selected variables (fruit consumption, vegetable consumption, dietary fat consumption and education) and revised meta-analyses conducted. Third, further adjustments are made for the biasing effects of misclassification of smoking status. The materials and methods section is therefore divided accordingly.

Study inclusion and exclusion criteria

Attention was restricted to epidemiological prospective or case-control studies published up to and including July 2015, which involved five or more lung cancers, and which provided relative risk (RR) estimates for never (or virtually never) smokers for one or more defined ETS exposure types or dose-related ETS indices. The “exposure types” compare subjects exposed and unexposed to ETS from seven different sources: Spouse; household; workplace; childhood; travel; social and other; and total, the final category including biochemical assessments of exposure. The “dose-related indices” concern ETS exposure in terms of amount smoked, duration of smoking and the number of smokers the subject was exposed to. ETS exposure from pipe/cigar only was ignored. Note that the term “relative risk” is taken to include estimates of it, such as the odds ratio or hazard ratio.

Studies using near equivalent definitions of “never smokers” were accepted when stricter definitions were unavailable, so never smokers could include occasional smokers, those with a minimal lifetime duration of smoking or number smoked, or ex-smokers who had quit at least 20 years previously.

Literature searches

Up until July 2015 potentially relevant papers were regularly sought from MEDLINE searches (using search terms “tobacco smoke pollution” and “lung neoplasm”), from files on smoking and health which were collected for many years within our company, and from references which were cited in the papers obtained. At the end of the process no paper examined cited a possibly relevant paper which had not been previously examined.

Study identification

Relevant papers were separated into studies, noting where there were multiple papers per study or multiple studies per paper, and any overlaps between studies. Each study was uniquely referenced by a ≤ 6 character code, based on the name of the principal author, with a suffix indicating where the same author had reported on multiple studies.

Data recorded

Data were entered on a study database, and also on a linked RR database. The structure and content of the databases are described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%201.pdf.

In brief, a study database record describes the study design, the available data and a previously described index of study quality[11]. Typically there are multiple records per study on the RR database, each record holding a detailed description of a specific comparison made and the corresponding RR and its 95%CI.

RR derivation

When available, adjusted RRs and CIs were entered. Unadjusted estimates were derived from the 2 × 2 table, using standard methodology[12], any differences between calculated and author-provided estimates being noted. Other methodologies were used where required to derive estimates, those more commonly used[13,14] being described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%201.pdf.

Identifying the RRs to enter

RRs were entered, if available, relating to various pre-defined combinations of type of lung cancer, index of smoking, confounders considered, and strata. The combinations are described in the following sections.

Type of lung cancer: Results were entered for overall lung cancer, squamous cell carcinoma and adenocarcinoma, or their nearest equivalents for which data were available.

Smoking indices: The intention was to enter RRs comparing subjects exposed and unexposed to the various indices of ETS defined above. Though RRs for exposure to smoking by the spouse should ideally be derived from data only for married subjects, we also allowed RRs from studies where unmarried subjects were included in the reference group. Similarly, RRs for workplace exposure could include non-working subjects. For the “household” and “childhood” categories, RRs were entered for all possible sources recorded by the studies, but for the “travel” and “social” categories, if more than one index of exposure was available, only that representing the greatest number of exposed subjects was entered. RRs were entered for all available timings of adult exposure, but for childhood, only RRs for the earliest exposure were entered. “Total” exposure was defined as exposure to two or more types of exposure, or biochemical assessment of overall exposure. For dose-related exposure indices, RRs were entered for each level of exposure relative to a common base level. RRs were entered, where available, using denominators representing both “no exposure to the specific type of ETS” and “no exposure to any ETS”.

Confounders: For case-control studies, we entered results adjusted for the most potential confounders available, and also adjusted for fewest. For prospective studies, we entered results adjusted for age and the most confounders, and for age and the fewest, and unadjusted results were entered only where there were no age-adjusted results. We describe these alternative RRs as “most-adjusted” and “least-adjusted”.

Strata: We only entered results stratified by sex or age. Combined sex results were only entered if results by sex were unavailable. We entered results for all ages and for separate age groups. Specifically for spousal exposure (or nearest equivalent - see “analyses conducted” below), where an adjusted RR was available only for combined sexes but numbers of cases and controls were given by sex, split-sex estimates were entered, assuming that the RR applied to each sex, with separate CIs estimated for males and females.

Meta-analyses

Analyses conducted: The series of meta-analyses conducted was pre-planned. For a given exposure type, a set of up to 20 analyses was conducted. Meta-analyses 1 and 2 used the overall data available, while meta-analyses 3 and 4 were separated by region (North America, Europe, Asia or other regions), with meta-analyses 1 and 3 using most-adjusted and 2 and 4 least-adjusted data. Analyses 5-20 were based on most-adjusted data only and studied variation by other factors, as described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%201.pdf.

The primary index of exposure used was “spousal smoking (or nearest equivalent)” where, for studies which provided no results for spousal exposure, results for household, total or both spousal/home and other exposure were chosen instead. This identified a single exposure definition for each study. For overall lung cancer, the full set of 20 meta-analyses was carried out restricted to females, and unrestricted on sex. Further meta-analyses for the principal index of exposure corresponded to meta-analyses 1 to 4 only. These included analyses for spousal smoking (or nearest equivalent) for males, spousal smoking (specifically) for females, males and unrestricted on sex, and analyses for spousal smoking (or nearest equivalent) for squamous cell carcinoma and for adenocarcinoma, each for females, males and unrestricted on sex.

Analyses for the other types of exposure were run only for overall lung cancer, without restriction on sex, and were equivalent to meta-analyses 1-4 only. The childhood and household exposure analyses were run using alternative indices, depending on the available data, as described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%201.pdf.

Selecting RRs for the meta-analyses: In selecting RRs to include we tried to include all relevant data once only. Where a study had multiple RRs, that used is chosen by an order of preference specific to the meta-analysis. Order of preference may be needed for exposure status and timing, and for the unexposed base. As RR definitions may be sex-specific, the RRs selected may differ by sex. Results for sexes combined are only considered in the absence of sex-specific results.

Conducting the meta-analyses for exposure indices: We conducted fixed-effect and random-effects meta-analysis of study-specific data for the various exposure indices studied as described elsewhere[13]. Heterogeneity was measured by H, the ratio of heterogeneity χ² to degrees of freedom. H relates to I² statistic[15] by I² = 100 (H - 1)/H. For all meta-analyses, results of publication bias tests using the Egger method[16] were also given.

Results are displayed in forest plots. Within each plot, studies are identified by their reference code, and listed in order of RR. Most of the plots are also grouped by region. The study estimates are shown both as numbers and in graphical form logarithmically. In the latter representation an RR is shown as a square, the area of which is proportional to its weight, its inverse-variance. Arrows warn if the CI goes outside the range of the plot. Random-effects estimates are also presented, overall or by region, shown by a diamond whose width indicates the 95%CI.

Carrying out meta-analyses for number of cigarettes smoked by the husband: The methods used are as described elsewhere[17], and are summarized here. The underlying model assumes that, when comparing two groups differing in exposure by dose d, log RR is estimated by βd. For each study, given data at each level of exposure consisting of the dose level, the number of cases, and the number of controls (or subjects at risk), β and its standard error (SEβ) are estimated by the method of Greenland and Longnecker[18], This can be applied to studies with only two levels (unexposed and exposed), and also to confounder-corrected RRs and 95%CIs, by estimating pseudo-counts using the method of Hamling[14]. Estimates of β and SEβ from each study are then meta-analysed as described above. The method of estimating midpoint doses for intervals such as 1-19 or 20+ cigarettes per day is as described previously[17].

The series of meta-analyses conducted for the estimates of β was similar to that for the exposure indices as described above.

Adjustment for bias due to confounding

The potential confounding variables considered (consumption of fruit, consumption of vegetables, consumption of dietary fat, and education) and the methods used to adjust for them are as described in a previous publication[19] and in an unpublished updated analysis conducted in 2006[20]. The methods are summarized briefly below.

Estimating the relationship between the four potential confounding variables considered and the risk of lung cancer in never smoking females: The database of studies relating lung cancer risk in never smokers to the four potential confounding variables used in 2001[19] and updated in 2006[20] was used, restricting attention to never smoking females. Exp(β), the increase in risk per dose unit (SDs for dietary variables, years for education), was estimated using methodology equivalent to that described in the previous section. Methods for assigning midpoint dose values for grouped dietary data (e.g., by quintiles), and for education groups were as before[19].

Estimating the differences in fruit, vegetable and fat consumption and in education associated with ETS exposure: The database of studies comparing the distribution of the four potential confounding variables set up in 2001[19] and updated in 2006[20] was used, with attention restricted to never smoking females. For each potential confounding variable, we estimated the difference, δ, in relation to marriage to, or living with, a smoker, in units of SDs for the dietary variables and years for education, using the methodology described earlier[19]. These study-specific estimates of δ were then combined using random-effects meta-analysis.

The basic method for confounder adjustment: As described earlier[19] we assume that the logarithm of L, the lung cancer risk, is linearly related to n explanatory factors x_i by:

log L = β_o^* + ∑_iⁿβ_i^*x_i (1)

Exp(β_o^*) is the background risk that is expected for zero exposure to each factor. Exp (β_i^*) is the multiplicative risk increase expected per unit increase in exposure to the i^th factor.

Should data relating lung cancer to the factors be available only univariately the relationship with each factor would be formulated as:

log L = β_o + β_ix_i (2)

where exp(β_i) is the RR for a unit dose increase associated with factor i that is not adjusted for the other risk factors.

The β_i and the β_i^* are related by the matrix equation:

B* = S^-1C^-1SB (3)

Here B* and B are the n × 1 vectors of and β_i, S the n x n standard deviation (SD) matrix, s_i, and C the n x n correlation matrix c_ij.

Given B, C and S, we can estimate B*. In our context, there are five factors. i = 1 represents ETS with I = 2…5 the three dietary variables and education. Thus exp(β_i) is the unadjusted risk increase for each unit of increase in ETS exposure, and exp(β_i^*) the adjusted risk increase. The joint confounding effect of the four variables is estimated as exp(β_i)/exp(β_i^*).

Relationship of the factors to lung cancer risk: Estimates of β_i are generally those described in the sub-section “carrying out meta-analyses for number smoked by the husband”. However, the basic method assumes that β_i is unadjusted for any of the four potential confounding variables. Where β_i is adjusted for one or more of the variables, we first back-corrected it in order to take out the effect of the adjustment as described earlier[19]. This back-correction procedure avoids double-adjustment for the same factor. Back-correction was also carried out in the following cases: For fruit consumption, where the RR estimate had already been adjusted for vitamin C; for dietary fat, where the RR estimate had already been adjusted for energy intake, for meat, or for cholesterol; and for education, where the RR estimate had already been adjusted for income, for socioeconomic status, or for ownership of a colour TV.

Standard deviations: We estimated s₁, the SD for ETS, directly for each study from the population data by level of exposure as described elsewhere[17]. s₂, s₃ and s₄, the SDs for the dietary variables are 1, since they are measured in units of SD. We took the SD for education as 2.435 years[20].

Correlations: If i = j, c_ij = 1. To quantify other correlations, we used the formula:

C_ij = δ_js₁/d₁(average)s_j (4)

Here δ_j is a common estimate of the difference in exposure to variable j for living with a smoker (see the sub-section “estimating the differences in fruit…”). s₁ and s_j are as described above, and d₁(average) is the mean ETS exposure for exposed never smokers. Where studies have more than one exposure level, we estimated d₁(average) by weighting on the number of exposed subjects.

To quantify the correlations between the potential confounding variables we used averaged data from seven databases, the five used in 2001[19] and two additional US databases (NHIS2000, NHANES III), as described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%202.pdf.

Adjustment for bias due to misclassification of smoking status

How the bias arises and what it depends on: Estimates of the RR of lung cancer in self-reported never smoking women associated with marriage to a smoker may be biased if a proportion of the women are actually current or ex-smokers. This bias arises because smokers marry smokers more often than is expected by chance. Misclassified smokers are therefore commoner among those married to a smoker. As shown by Lee and Forey[21], the bias depends mainly on the rate of misclassification, the active smoking risk, the degree to which smoking by spouses is concordant, and the proportions of smokers among subjects and their spouses.

Correction method used: We use the Lee and Forey[21] method for bias correction, assuming joint effects of active smoking and ETS exposure are additive, and the published extension of the method[22].

Concordance ratio: The concordance ratio is defined as the odds of the husband smoking if the wife ever smoked divided by the odds if the wife never smoked. From an earlier review[23] we used an estimate of 3.0.

Study-specific data on active smoking RRs: For each study, estimates were made of the active smoking RR, derived if possible from the source paper itself or another paper using the same study population. Otherwise they were derived from studies in that country, from estimates presented by the EPA[24] or by other methods, as described in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%202.pdf.

Misclassification rates: Misclassified smokers have a lower lung cancer risk than non-misclassified smokers. To take this into account, we followed precedent[21,22] in carrying out the misclassification correction on the basis that those who deny smoking have the same risk as those who admit it, but using lower misclassification rates (10% for Asia, and 2.5% elsewhere) than are observed. Support for the use of these rates is provided elsewhere[21-23,25,26].

Application of the method: RRs for spousal smoking and for amount smoked by the spouse, were calculated: (1) with no adjustment for confounding or correction for misclassification; (2) with adjustment for confounding and no correction for misclassification; and (3) with adjustment for confounding and correction for misclassification.

RESULTS

Studies identified

There were 102 studies which met the inclusion criteria. Some studies were noted to have overlaps with other studies. However, as all overlaps were minor and could not be disentangled, it was decided to ignore them. Tables 1 and 2 give study details including reference(s), location, design, dates, numbers of cases in never smokers, definition of never smoking, ETS exposure measures considered, adjustment variables used, extent of histological confirmation of cases, whether results are available by histological type, and availability of dose-response data. http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%203.pdf describes why other publications which could be thought possibly relevant are not considered in our analyses.

Table 1 Selected details of the 102 studies in publication date order.

Study ref	Main ref	Other ref	Location	Study design1	Study dates2			Total cases	Never smoker criteria3
GARFI1	[75]	[104]	United States	P	1959 -	1960,	1971	153	Dur 6 mo
CHAN	[105]	[106,107]	Hong Kong	CC	1976 -	1977		84
CORREA	[67]		United States	CC	1979 -	1982		32
TRICHO	[27]	[108,109]	Greece	CC	1978 -	1982		77
BUFFLE	[110]		United States	CC	1976 -	1980		52
HIRAYA	[28]	[111-120]	Japan	P	1965 -	1966,	1981	264
KABAT1	[53]		United States	CC	1971 -	1980		76
GARFI2	[121]	[122]	United States	CC	1971 -	1981		134
LAMW	[29]		Hong Kong	CC	1981 -	1984		75
WU	[123]		United States	CC	1981 -	1982		31
AKIBA	[68]		Japan	CC	1971 -	1980		113
LEE	[124]		United Kingdom	CC	1979 -	1982		47
BROWN1	[125]		United States	CC	1979 -	1982		19
GAO	[126]		China	CC	1984 -	1986		246
HUMBL1	[127]		United States	CC	1980 -	1984		28	Dur 6 mo
KOO	[128]	[129-133]	Hong Kong	CC	1981 -	1983		88	N20
LAMT	[30]		Hong Kong	CC	1983 -	1986		199	Occ (1 yr)
PERSHA	[134]	[135]	Sweden	CC	1961 -	1980		77
BUTLER	[136]		United States	P	1976 -	1976,	1982	8
GENG	[31]	[137]	China	CC	1983 -	1983		54
INOUE	[138]		Japan	CC	1973 -	1983		28
SHIMIZ	[70]		Japan	CC	1982 -	1985		90
CHOI	[139]		South Korea	CC	1985 -	1988		88
HOLE	[140]	[141]	Scotland	P	1972 -	1976,	1985	9
SCHOEN	[142]		United States	CC	1982 -	1983		116
SVENSS	[143]	[144]	Sweden	CC	1983 -	1985		34	Occ
JANERI	[72]	[145]	United States	CC	1982 -	1985		191	N100
KALAND	[32]		Greece	CC	1987 -	1989		91
SOBUE	[146]	[147]	Japan	CC	1986 -	1988		144
WU-WIL	[46]		China	CC	1985 -	1987		417
LIUZ	[148]	[149,150]	China	CC	1985 -	1986		54	1sm
BROWN2	[151]	[152-155]	United States	CC	1986 -	1991		432
STOCKW	[58]	[156]	United States	CC	1987 -	1991		210	N100, Dur 6 mo
DU	[52]	[157-159]	China	CC	1986 -	1986		75
LIUQ	[160]		China	CC	1983 -	1984		38
FONTHA	[33]	[161-164]	United States	CC	1986 -	1988		653	N100, Dur 6 mo
LAYARD	[165]		United States	CC	1986 -	1986		60
DEWAAR	[166]	[167]	Netherlands	CC	1977 -	1991		23	Cot
KABAT2	[168]	[169,170]	United States	CC	1983 -	1990		110
SCHWAR	[57]		United States	CC	1984 -	1987		257
SUN	[60]		China	CC	NA			230
WANGS	[34]		China	CC	NA			82
WANGT	[171]		China	CC	1992 -	1994		135
CARDEN	[73]	[172,173]	United States	P	1982 -	1982,	1989	362
ZHENG	[35]		China	CC	1990 -	1993		94	Non
AUVINE	[174]		Finland	CC	1986 -	1992		44
BOFFET	[66]	[175-180]	West Europe	CC	1988 -	1994		640	N400
SHEN	[181]	[182-185]	China	CC	1993 -	1993		70
ZARIDZ	[36]	[186-188]	Russia	CC	1991 -	1993		189
BOFFE2	[189]		Europe	CC	1994 -	1996		70	N400
JEE	[190]		South Korea	P	1992 -	1994,	1997	79	Occ, Dur 1 yr
RAPITI	[61]		India	CC	1991 -	1992		58	N400
SPEIZE	[191]		United States	P	1982 -	1982,	1992	35
ZHONG	[54]		China	CC	1992 -	1994		504	Dur 6 mo
LEECH	[37]	[192-194]	Taiwan	CC	1992 -	1998		268	N400
MALATS	[195]	[196]	Europe/Brazil	CC	NA			122	Occ
WANGL	[62]	[197]	China	CC	1994 -	1998		233	Dur 6 mo
JOHNSO	[198]	[199-201]	Canada	CC	1994 -	1997		71	N100
LAGARD	[202]	[203]	Sweden	CC	1980 -	1995		433	Occ (1 yr)
NISHIN	[204]		Japan	P	1984 -	1984,	1992	24
OHNO	[205]		Japan	CC	NA			191	N365
RACHTA	[63]	[206]	Poland	CC	1991 -	1997		54	Dur 6 mo
ENSTRO	[207]	[208]	United States	P	1959 -	1960,	1998	256
ZATLOU	[64]	[209,210]	Czech Republic	CC	1998 -	2002		84	N100
IARCKR	[1]	[180]	Germany	CC	1990 -	1996		123	Dur 6 mo, N400
MCGHEE	[211]		Hong Kong	CC	1998 -	1998		324
EPICA	[212]	[213-216]	Western Europe	P	1993 -	1998,	2000	59
FANG	[38]	[216]	China	CC	2001 -	2004		157
FRANCO	[71]	[216]	Mexico	CC	2000 -	2002		94
GORLOV	[55]	[217-222]	United States	CC	1995 -	2003		193	N100
NEUBER	[49]		United States	CC	1994 -	1997		56	N100, Dur 6 mo
RYLAND	[223]		Sweden	CC	1989 -	1994		49
WEN	[56]	[224-227]	China	P	1997 -	2000,	2004	106
YU	[228]	[228-230]	Hong Kong	CC	2002 -	2004		213	N400, Dur 1 yr
ZEKA	[59]		East Europe, United Kingdom	CC	1998 -	2002		223	N100
HILL1	[231]		New Zealand	P	1981 -	1981,	1984	147	Occ
HILL2	[231]		New Zealand	P	1996 -	1996,	1999	234	Occ
LOPEZC	[232]		Spain	CC	2000 -	2005		36	N100
ASOMAN	[233]		United States	CC	1992 -	NA		138	Occ (1 yr)
GALLEG	[51]		Mexico	CC	2003 -	2007		32
KURAHA	[234]		Japan	P	1990 -	1993,	2004	109
PANDEY	[50]		Nepal	CC	NA			268
YANG	[39]	[65,221,235,236]	United States	CC	1997 -	2008		297	N100
OLIVOM	[65]	[237]	United States	CC	NA			45	N100
TSE	[238]	[239]	China	CC	2004 -	2006		132	N400, Dur 1 yr
LIANG	[40]		China	CC	2004 -	2007		226
BRENNE	[47]		Canada	CC	1997 -	2002		156	N100
JIANG	[41]		China	CC	2009 -	2009		145
EPICC	[240]	[212]	Western Europe	P	1992 -	1998,	2006	98
KIYOHA	[241]		Japan	CC	1996 -	2008		153
HE	[242]	[243]	China	P	1976 -	1994,	2011	16	N100
LIM	[74]	[244-246]	China	CC	1996 -	2008		433	Occ (1 yr)
LIN	[42]		China	CC	2006 -	2010		226
FERREC	[247]		Chile	CC	2007 -	2010		59
ALZOUG	[48]	[248,249]	Canada	CC	1996 -	2000		44	N100
GELAC	[43]		Taiwan	CC	2002 -	2009		1540	Occ
MASJED	[44]	[250]	Iran	CC	2002 -	2005		81	Dur 6 mo
REN	[251]		China	CC	2002 -	2012		764
SEKI	[252]		Japan	CC	1997 -	2009		431
WHIOS	[253]	[254]	United States	P	1993 -	1998,	2009	200
ILCCO	[45]	[69]	International	CC	1984 -	2014		2504	N100
TORRES	[255]	[256]	Spain	CC	2011 -	2013		192	N100

¹Study design is coded as P: Prospective; CC: Case control;

²Study dates are given as Start year, End year, Final follow-up year (prospective studies only);

³Inclusion of “near equivalents” to never smokers, coded as Dur: Includes those who smoked up to a number of months (mo) or years (yr); N: Includes those who smoked up to a number of cigarettes in their lifetime; Occ: Includes occasional smokers; Occ (1 yr): Includes those who smoked occasionally for up to 1 year; Non: Described as “non-smokers” and assumed from context to mean never smokers; 1sm: Study included 1 smoker; Cot: Excluded self-reported never smokers with urinary cotinine > 100 ng/mg.

Figure 1 Forest plots for smoking by husband, by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1. In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

Table 2 Further details of the 102 studies.

Ref.	ETS exposures1	Adjustment variables used2	Extent (%) of histological confirmation	Results by histological type	Dose-response results available3
GARFI1	s	7	NA		Yes
CHAN	s	0	80
CORREA	s c4	1	97		Yes
TRICHO	s	0	27	Yes	Yes
BUFFLE	h	0	100		Yes
HIRAYA	s	2	NA		Yes
KABAT1	s h w	4	100
GARFI2	s h w c o tot	4	100	Yes	Yes
LAMW	s tot	1	100	Yes
WU	s w c tot4	2	100	Yes
AKIBA	s c	6	53		Yes
LEE	s h w tr o tot	3	38	Yes
BROWN1	tot	3	100	Yes
GAO	s h c tot	2	43		Yes
HUMBL1	s	2	100		Yes
KOO	s h w c tot	5	97	Yes	Yes
LAMT	s	0	100	Yes	Yes
PERSHA	s c	2	83	Yes	Yes
BUTLER	s	2	100
GENG	s	0	85		Yes
INOUE	s	3	NA		Yes
SHIMIZ	s h w	3	100
CHOI	s	0	100		Yes
HOLE	h	2	NA		Yes
SCHOEN	s	6	100
SVENSS	h c tot	1	70
JANERI	s h w c o	3	100	Yes	Yes
KALAND	s h w	5	48	Yes	Yes
SOBUE	s h c	3	100
WU-WIL	s h w c	5	42
LIUZ	h	3	17
BROWN2	s h w c	2	76	Yes	Yes
STOCKW	s h w4 c o4 tot	3	100	Yes	Yes
DU	s	2	NA		Yes
LIUQ	s	3	32		Yes
FONTHA	s h w c o tot	10	100	Yes	Yes
LAYARD	s	3	NA		Yes
DEWAAR	tot	0	71		Yes
KABAT2	s h w c tr o tot	6	100		Yes
SCHWAR	h w	2	100
SUN	s h w c hw tot	2	100	Yes
WANGS	tot	0	100
WANGT	s w c	1	57		Yes
CARDEN	s h w o tot	8	NA		Yes
ZHENG	h	2	82	Yes	Yes
AUVINE	tot	1	NA
BOFFET	s h w c tr o tot	7	96	Yes	Yes
SHEN	tot	9	100	Yes	Yes
ZARIDZ	s h w c	2	100	Yes	Yes
BOFFE2	s w c tot	5	100	Yes	Yes
JEE	s	5	0		Yes
RAPITI	s w c tr	3	100	Yes	Yes
SPEIZE	tot	1	100
ZHONG	s h w c tot	7	57	Yes	Yes
LEECH	s h w c tot	7	100		Yes
MALATS	s tot	2	100		Yes
WANGL	h c	6	32		Yes
JOHNSO	h w c tot	4	100		Yes
LAGARD	h tot	6	100
NISHIN	s h	7	NA
OHNO	s w c tr o tot	2	100		Yes
RACHTA	c	21	100	Yes
ENSTRO	s	8	NA
ZATLOU	c tot	3	100	Yes
IARCKR	s w c	2	100		Yes
MCGHEE	h	2	0		Yes
EPICA	h w tot	7	NA
FANG	tot	8	100		Yes
FRANCO	h	2	100
GORLOV	h w hw tot	4	100
NEUBER	tot4	3	100
RYLAND	h w	3	98
WEN	s w c tot	9	NA		Yes
YU	tot	20	100		Yes
ZEKA	w	4	NA		Yes
HILL1	h	9	NA
HILL2	h	9	NA
LOPEZC	tot	0	100
ASOMAN	h w o	0	100
GALLEG	tot	0	100
KURAHA	s w c hw tot	5	90	Yes	Yes
PANDEY	c tot	0	NA		Yes
YANG	c tot	5	NA
OLIVOM	c	4	100	Yes
TSE	h w tot	9	100	Yes	Yes
LIANG	c tot	0	100
BRENNE	h w c tot	3	100		Yes
JIANG	tot	17	100	Yes
EPICC	c	10	NA		Yes
KIYOHA	s	0	100
HE	tot	9	88		Yes
LIM	h	0	96
LIN	tot	9	100		Yes
FERREC	c tot4	3	72
ALZOUG	s h w c tot	3	NA	Yes	Yes
GELAC	s h w hw tot	6	100		Yes
MASJED	s h w o tot	4	100	Yes
REN	tot	5	100
SEKI	s	7	94	Yes
WHIOS	h4 w4 c4 tot	0	0		4
ILCCO	h w c tot	3	100	Yes	Yes
TORRES	h	2	99		Yes

¹ETS exposure measures reported, coded as s: Spousal; h: Household; w: Work; c: Childhood; hw: Exposure at both home and work; tr: Travel; o: Social/other; tot: Total exposure. Codes marked 4 represent exposures for which the only result presented is a statement that no association was found;

²Number of factors adjusted for, excluding sex;

³Coded as yes: Dose response result presented;

⁴The only dose response result presented is a statement that no dose response was found. NA: Not available.

Figure 2 Forest plots for smoking by wife, by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1. In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RRs shown with a - are calculated using a 0.5 addition to each cell, due to a zero in the 2 x 2 table. RR: Relative risk.

Of the 102 studies, 31 were conducted in North America (including 26 in United States), 45 in Asia (including 23 in China, 10 in Japan and 6 in Hong Kong), 21 in Europe (4 in Sweden being the most for any country), and 5 in other locations (including two international studies).

Eighty-five studies were of case-control design and 17 were prospective. Twenty-six studies were published in 1981-1989, 28 in 1990-1999, 32 in 2000-2009 and 16 in or after 2010.

In general, the total number of cases per study was small, with 20 studies based on less than 50 cases, and 29 considering 50-99. Twenty-four studies examined 100-199 cases, 18 200-399 cases, with only 11 based on 400 or more cases.

The most commonly studied index was smoking by the spouse, considered by 55 studies. Smoking by a cohabitant was considered by 47 studies, workplace smoking by 40, and childhood exposure by 41. Travel and social exposures were considered by 5 and 11 studies respectively, and total exposure by 51.

Effect estimates

In what follows, meta-analysis RRs referred to in the text, tables and figures are based on “most-adjusted” estimates, meta-analysis RRs based on “least-adjusted” estimates usually being very similar. The results highlighted are drawn from more detailed analyses for all the exposure indices made available in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%204.pdf, which also shows the “preferences” used in each analysis. This includes some analyses based on “least-adjusted” estimates, and also gives estimates for each individual study included in an analysis. Significance is taken to be at P < 0.05 unless otherwise stated. RRs and 95%CIs are normally shown to 2 decimal places. Exceptionally, they are shown to 3 decimal places for the analyses investigating bias due to confounding and misclassification, to show the effects of adjustment and correction more clearly.

Smoking by the spouse

Ninety-three studies provided results relating lung cancer in women to husband’s smoking (or nearest equivalent), with 19[27-45] giving significantly raised RRs and 51 non-significantly raised RRs. In contrast 18 studies showed a negative relationship, significant in three studies[46-48]. Five studies gave a RR of 1.00. Two studies could not be included in the meta-analysis, one study[49] reporting no significant effect of passive smoking but giving no further details, with another[50] only giving an odds ratio of 2.2 (1.4-3.7) for greater than 40 smoker-years exposure to passive smoking. There was marked heterogeneity (P < 0.001) between the individual study estimates. However, fixed-effect (1.19, 95%CI: 1.14-1.24) and random-effects estimates (1.22, 1.14-1.31) were similar. Based on the Egger test there was no clear evidence of publication bias (0.05 < P < 0.1). Further analyses of these data are given in the section “smoking by the husband - detailed analyses” below.

Forty-five studies considered smoking by the wife (or nearest equivalent). Twenty-eight RRs were raised, three[28,39,51] significantly. Fifteen studies reported a negative association, significant in one study[47]. One study found no association. One study[52] reported ETS was not statistically associated with lung cancer, but gave no further details and could not be included in the meta-analysis. Heterogeneity between studies was not significant, and fixed-effect (1.15, 1.03-1.28) and random-effects estimates (1.14, 1.01-1.29) were similar. There was no evidence of publication bias (P > 0.1).

Results by sex, separated by region, are given as forest plots in Figure 1 (husband smoking) and Figure 2 (wife smoking).

Further meta-analyses were carried out on results for smoking by either the husband or the wife (or nearest equivalent), based on 119 RR estimates. Details are given in Table 3, along with estimates split by various other factors. Overall, a fixed-effect RR (1.18, 1.14-1.23) and a random-effects RR (1.21, 1.14-1.29) were estimated, with marked heterogeneity between studies (P < 0.001). When the studies were examined according to various factors, there was evidence of heterogeneity between factor levels for publication date (P < 0.01), study size (P < 0.01) and age adjustment (P < 0.05), with the largest RRs seen for early (1981-1989) studies, small studies (1-49 cases) and estimates unadjusted for age. There was no significant heterogeneity by location, study type, reporting of dose-response results, or use of spouse as the index of exposure. There was no clear evidence of publication bias (0.05 < P < 0.1).

Table 3 Meta-analyses of smoking by the spouse (or nearest equivalent).

Estimates considered	No. of estimates	Relative risk (95% confidence limits)		Heterogeneity1
		Fixed-effects meta-analysis	Random-effects meta-analysis
All	119	1.18 (1.14-1.23)	1.21 (1.14-1.29)	+++
N America	38	1.08 (1.00-1.17)	1.09 (0.95-1.26)	+++
Europe	22	1.15 (1.03-1.28)	1.16 (1.00-1.35)	(+)
Asia	54	1.24 (1.17-1.32)	1.31 (1.20-1.44)	+++
Asia - Japan	13	1.26 (1.11-1.45)	1.26 (1.11-1.45)	NS
Asia - Hong Kong	8	1.32 (1.12-1.57)	1.31 (1.06-1.63)	NS
Asia - China	23	1.16 (1.06-1.27)	1.29 (1.08-1.54)	+++
Asia - Other	10	1.34 (1.19-1.51)	1.37 (1.19-1.57)	NS
Heterogeneity between Asian countries				NS
Other continents	5	1.20 (1.09-1.33)	1.20 (1.09-1.33)	NS
Heterogeneity between continents				(+)
Published in 1981-1989	34	1.38 (1.24-1.54)	1.38 (1.24-1.54)	NS
Published in 1990-1999	33	1.09 (1.01-1.17)	1.15 (1.02-1.28)	++
Published in 2000-2009	34	1.22 (1.12-1.33)	1.21 (1.08-1.36)	+
Published in 2010 onwards	18	1.17 (1.10-1.26)	1.13 (0.94-1.36)	+++
Heterogeneity by publication date				++
1-49 cases	23	1.44 (1.14-1.81)	1.47 (1.15-1.88)	NS
50-99	31	1.30 (1.14-1.47)	1.27 (1.08-1.50)	+
100-199	29	1.09 (1.00-1.19)	1.10 (0.96-1.26)	+++
200-399	22	1.33 (1.21-1.46)	1.32 (1.16-1.50)	+
400+	14	1.14 (1.07-1.20)	1.13 (1.02-1.25)	++
Heterogeneity by study size				++
Case-control	97	1.18 (1.13-1.23)	1.22 (1.13-1.31)	+++
Prospective	22	1.18 (1.05-1.33)	1.18 (1.05-1.33)	NS
Heterogeneity by study type				NS
Not age adjusted	21	1.34 (1.19-1.50)	1.42 (1.18-1.71)	+
Age adjusted	98	1.16 (1.11-1.21)	1.18 (1.10-1.26)	+++
Heterogeneity by age adjustment				NS
Dose-response results not reported	46	1.13 (1.06-1.21)	1.18 (1.06-1.31)	++
Only no dose-response stated	2	0.95 (0.60-1.50)	0.95 (0.60-1.50)	NS
Dose-response results reported	71	1.21 (1.15-1.28)	1.24 (1.14-1.35)	+++
Heterogeneity by dose response reporting				NS
Spouse the index	71	1.18 (1.11-1.24)	1.21 (1.12-1.31)	++
Spouse not the index	48	1.19 (1.12-1.27)	1.20 (1.07-1.35)	+++
Heterogeneity by index definition				NS

¹Significance levels indicated by +++ P < 0.001, ++ P < 0.01, + P < 0.05, (+) P < 0.1 for heterogeneity within level and for heterogeneity between level. NS: Not significant, P≥ 0.1.

Results for smoking by the spouse (or nearest equivalent) were also examined by histological type of cancer, with Figure 3 (squamous cell carcinoma) and Figure 4 (adenocarcinoma) showing forest plots by region. The analysis of squamous cell carcinoma, based on 24 RR estimates, showed a significant (P < 0.001) positive association and heterogeneity (P < 0.001), overall estimates being 1.41 (1.24-1.59, fixed-effect) and 1.44 (1.15-1.80, random-effects). No significant variation by region was seen. For adenocarcinoma, the 30 RR estimates were again heterogeneous (P < 0.01), with the meta-analysis showing significantly raised RRs, of 1.23 (1.15-1.32, fixed-effect) and 1.33 (1.17-1.51, random-effects). The heterogeneity was partly due to differences (P < 0.001) by region, with little increase seen in North American and European studies (random-effects RRs 1.08, 0.96-1.22 for North America; 1.11, 0.82-1.49 for Europe), but a clear increase for Asia (random-effects RR 1.70, 1.35-2.15).

Figure 3 Forest plots for squamous cell carcinoma and spousal smoking, by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight RRs shown with a - are calculated using a 0.5 addition to each cell, due to a zero in the 2 x 2 table. RR: Relative risk.

Figure 4 Forest plots for adenocarcinoma and spousal smoking, by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

Workplace ETS exposure

For lung cancer and workplace ETS exposure, 47 RR estimates were available (Figure 5). Of these, 37 were raised, with estimates from six studies[33,45,53-56] being significant, and another[57] of borderline significance. This contrasted with nine studies, where RRs were non-significantly below 1.00, and one showing no association. Two other studies[58,59], neither of which reported an association, could not be included in the meta-analysis, due to providing insufficient detail. Overall, there was a significant positive relationship, whether based on fixed-effect (1.21, 1.14-1.28) or random-effects RRs (1.22, 1.15-1.30). There was no evidence of heterogeneity or publication bias. Studies conducted in North America (1.21, 1.08-1.37), Europe (1.18, 1.01-1.39) and Asia (1.33, 1.20-1.47) all showed a significantly increased random-effects RR.

Figure 5 Forest plots for workplace environmental tobacco smoke exposure by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RRs shown with a - are calculated using a 0.5 addition to each cell, due to a zero in the 2 x 2 table. RR: Relative risk.

Childhood ETS exposure

Results for childhood ETS exposure are given, by region, in Figure 6, with further meta-analyses given in Table 4. For childhood exposure from any cohabitant, 41 RR estimates were available. Of these, 21 were raised, eight significantly[37,39,60-65]. In contrast 18 RR estimates were below 1.00, one[66] significantly so, while two were equal to 1.00. In addition, three studies[67-69] found no relationship but provided insufficient detail for inclusion in the meta-analysis. Although meta-analysis suggested a positive relationship with the risk of lung cancer, this only just reached statistical significance (fixed-effect RR = 1.08, 1.01-1.15; random-effects RR = 1.15, 1.02-1.29). There was significant heterogeneity between the studies (P < 0.001), and heterogeneity between the continents (P < 0.05), with a significant increase seen in Asia (random-effects RR = 1.31, 1.02-1.67), but not in North America (RR = 1.06, 0.89-1.28) or Europe (RR = 1.02, 0.81-1.29).

Table 4 Meta-analyses of childhood environmental tobacco smoke exposure.

Estimates considered	No. of estimates	Relative risk (95% confidence limits)		Heterogeneity1
		Fixed-effects meta-analysis	Random-effects meta-analysis
From any cohabitant	41	1.08 (1.01-1.15)	1.15 (1.02-1.29)	+++
N America	13	1.00 (0.89-1.13)	1.06 (0.89-1.28)	+
Europe	11	0.94 (0.81-1.08)	1.02 (0.81-1.29)	+
Asia	15	1.26 (1.11-1.42)	1.31 (1.02-1.67)	+++
Other	2	1.10 (0.95-1.28)	1.13 (0.89-1.45)	NS
Heterogeneity between continents				+
From mother specifically	9	0.96 (0.77-1.20)	0.98 (0.77-1.25)	NS
From father specifically	11	0.90 (0.80-1.00)	1.00 (0.78-1.29)	+++
From parents specifically	8	0.78 (0.64-0.94)	0.78 (0.64-0.94)	NS

¹Significance levels indicated by +++P < 0.001, ++P < 0.01, +P < 0.05, (+) P < 0.1 for heterogeneity within level and for heterogeneity between level. NS: Not significant, P≥ 0.1.

Figure 6 Forest plots for childhood environmental tobacco smoke exposure by region. Estimates of the random-effects RR and its 95%CI are shown separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

Based on nine RR estimates, meta-analysis showed no evidence of any relationship specifically with maternal smoking in childhood, with the fixed-effect estimate 0.96 (0.77-1.20) and the random-effects estimate 0.98 (0.77-1.25). There was also no association specifically with paternal smoking in childhood (fixed-effect model 0.90, 0.80-1.00; random-effects model 1.00, 0.78-1.29), based on 11 estimates. However, meta-analysis of the eight estimates specifically considering parental smoking during childhood gave a reduced RR (0.78, 0.64-0.94 for both models). There was no significant heterogeneity for maternal smoking or parental smoking, but there was for paternal smoking (P < 0.001) due to an atypically high estimate of 12.64 (4.89-32.68) for females in one study[61].

None of the analyses of childhood exposure showed any significant evidence of publication bias.

Household ETS exposure

A total of 58 RR estimates were available for household ETS exposure from any source, as shown in Figure 7. Thirty-six RRs were above 1.00, statistically significant in six studies[35,43,45,60,70,71]. Twenty-one non-significantly negative RRs were also reported, while one study found no association. Overall RRs were 1.13 (1.07-1.19, fixed-effect) and 1.11 (1.05-1.18, random-effects). There was marked heterogeneity (P < 0.001) between the estimates, but no significant variation by study location, or evidence of publication bias.

Figure 7 Forest plot for household environmental tobacco smoke exposure. Estimates of the random-effects RR and its 95%CI are shown, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Overall estimates of RRs, 95%CIs and weights are also shown. Studies are identified by the study reference code shown in Table 1. In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

Restricting attention to sources of ETS other than the spouse, only 13 RRs were available, and the overall RR, although raised, was not significant (1.04, 0.89-1.21, fixed-effect) or (1.12, 0.87-1.44, random-effects).

ETS exposure during travel

Figure 8 shows the eight RRs for ETS exposure during travel. Six were above 1.00, and two were below 1.00. Only one estimate[61] was significant, and its high RR of 5.20 was the major contributor to the significant (P < 0.05) heterogeneity. Fixed effect meta-analysis gave a RR which was just significant (1.24, 1.01-1.53), but the random-effects RR was not (1.34, 0.94-1.93).

Figure 8 Forest plot for exposure to environmental tobacco smoke during travel. Estimates of the random-effects RR and its 95%CI are shown, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Overall estimates of RRs, 95%CIs and weights are also shown. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

ETS exposure in social situations

Analysis of the relationship of lung cancer to ETS exposure in social situations was based on 15 RR estimates, shown in Figure 9. One[33] study provided a significantly increased RR, with seven studies giving non-significantly raised estimates. Seven RRs were below 1.00, significantly so in two[72,73] studies. Overall, there was no evidence of an increased risk, for either fixed-effect (1.03, 0.92-1.16) or random-effects RRs (1.01, 0.82-1.24).

Figure 9 Forest plot for social environmental tobacco smoke exposure. Estimates of the random-effects RR and its 95%CI are shown, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Overall estimates of RRs, 95%CIs and weights are also shown. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RRs shown with a - are calculated using a 0.5 addition to each cell, due to a zero in the 2 x 2 table. RR: Relative risk.

Total ETS exposure

The 48 RRs for total ETS exposure are shown, by region, in Figure 10. Thirty-eight were above 1.00, significantly so for 12 studies[29,34,37-39,41-43,45,51,55,60]. Eight non-significantly reduced RRs were also reported, while two studies reported RRs of 1.00. Although there was marked heterogeneity (P < 0.001), fixed-effect RRs (1.30, 1.22-1.38), and random-effects RRs were quite similar (1.31, 1.19-1.45). Heterogeneity between the continents was statistically significant (P < 0.01), with random-effects RRs higher for Asia (1.51, 1.31-1.74), than for North America (1.22, 0.96-1.55) or Europe (1.09, 0.91-1.31). There was no evidence (P > 0.1) of publication bias.

Figure 10 Forest plots for total environmental tobacco smoke exposure, by region. Estimates of the random-effects RR and its 95%CI are shown, separately by region, sorted in increasing order of RR. These are shown numerically, and also graphically on a logarithmic scale. Weights (inverse-variance of log RR) are also shown, expressed as a percentage of the weight for all studies combined. Estimates of RRs, 95%CIs and weights are also shown for each region combined and overall. Studies are identified by the study reference code shown in Table 1, with sex identified by m (male), f (female) or c (sexes combined). In the graphical representation, individual RRs are indicated by a solid square, with the area of the square proportional to the weight. RR: Relative risk.

Smoking by the husband - detailed analyses

Smoking by the husband (or nearest equivalent) is now considered in more detail, with results presented both for overall exposure and per 10 cigarettes per day smoked by the husband. A fuller report which includes adjustment for confounding and for misclassification of exposure, is available in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%202.pdf, with only the main findings presented here.

For overall exposure, the RRs considered are those shown in Figure 1 and briefly referred to in the section “smoking by the spouse”. As noted there, combining estimates from 93 studies gave (RR = 1.19, 95%CI: 1.14-1.24, fixed-effects) and (RR =1.22, 95%CI: 1.14-1.31, random-effects).

Of the 93 studies, 29 were in North America, 18 in Europe, 26 in China or Hong Kong, 18 in the rest of Asia, and two in New Zealand. One Asian study[74] was of Chinese women in Singapore, and has been included in the subset of China studies. As the studies in New Zealand were principally of people of European descent, they have been included in the European subset of studies. One of the studies[45] was international, but due to a high proportion of Asian subjects has been included in the Rest of Asia subset.

The first study appeared in 1981[75], a further 25, 27, 26 and 14 being published in, respectively, 1982-89, 1990-99, 2000-09 and 2010-2014. Sixteen were prospective (cohort) studies and 77 case-control. Twenty-two studies involved less than 50 cases in lifelong non-smokers, and nine over 400 cases.

Nine studies adjusted for fruit consumption, 11 for vegetables, and 4 for dietary fat. Less than half (32 studies) adjusted for an index of education.

Twenty-four of the studies provided data on lung cancer risk by amount smoked by the husband specifically, with the remainder only providing results for overall exposure. Table 1 of http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%202.pdf gives the data used for each study and the fitted estimates of β and SEβ. Based on these data, it was estimated that each 10 cigarettes per day smoked by the husband multiplied risk by an estimated 1.09 (95%CI: 1.07-1.11) based on a fixed-effects analysis and 1.10 (1.07-1.14) using a random-effects analysis.

In order to adjust for the uncontrolled effects of confounding by diet (by fruit, vegetables and dietary fat) and education, summary estimates were required of the relationships of the four potential factors to both risk of lung cancer and ETS exposure, and of the correlations between the four factors. The estimates used are presented in Table 5, and show that, in non-smoking females, both risk of lung cancer and ETS exposure at home are associated with reduced fruit and vegetable consumption and education, and increased dietary fat consumption. All these associations are significant at least at P < 0.05, and for education at P < 0.001, with the data based on analysis of results from at least 10 studies (with one exception - dietary fat and lung cancer, based on 6 studies). Table 5 also shows the inter-correlations between the four confounding variables, based on combined estimates from seven studies. These show that fruit and vegetable consumption are strongly correlated with each other (P < 0.01). Other correlations are weaker and not significant at P < 0.05.

Table 5 Estimates used when adjusting for potential confounding effects¹.

	Statistic	Fruit consumption	Vegetable consumption	Dietary fat consumption	Education
Lung cancer risk	N studies	14	16	6	12
	RR2	0.86	0.88	1.22	0.91
	(95%CI)	(0.78-0.96)8	(0.80-0.97)8	(1.09-1.36)8	(0.88-0.95)6
	per	SD	SD	SD	Year3
ETS exposure	N studies	11	16	12	13
at home	Difference24	-0.0737	-0.0568	0.1317	-0.5346
	(SE)	-0.02	-0.021	-0.032	-0.063
	unit	SD	SD	SD	Year3
Correlations5	Fruit consumption	1	+0.3147	-0.104NS	+0.143NS
	Vegetable consumption		1	-0.054NS	-0.1309
	Dietary fat consumption			1	-0.039NS
	Education				1

Note: P values are indicated by

⁶P < 0.001,

⁷P < 0.01,

⁸P < 0.05,

⁹P < 0.1, or

^NSP≥ 0.1. ¹All data are for lifelong non-smoking females;

²Based on random-effects meta-analysis;

³The SD for education was taken as 2.435 years based on six studies;

⁴Difference in level of confounder between those exposed and unexposed to ETS at home;

⁵Based on seven studies, using unweighted means.

As described in the methods, we used misclassification rates of 10% for Asian studies and 2.5% elsewhere, these rates accounting for the lower rates of lung cancer seen among misclassified smokers than among non-misclassified smokers.

Table 6 presents results of analyses adjusting for confounding and misclassification based on RRs for smoking by the husband, while Table 7 similarly presents results based on RRs per 10 cigarettes smoked by the husband. Each table presents three sets of results: (1) unadjusted; (2) adjusted for confounders; and (3) adjusted for confounders and corrected for smoking misclassification. They each give overall estimates and results subdivided by various aspects of the studies considered.

Table 6 Adjusted/corrected analyses: Husband smoking¹.

Studies	n	Unadjusted for confoundingUncorrected for misclassificationRR (95%CI)	Adjusted for confounding2Uncorrected for misclassificationRR (95%CI)	Adjusted for confounding2Corrected for misclassification3RR (95%CI)
All	93	1.219 (1.138-1.305)	1.139 (1.062-1.221)	1.077 (0.999-1.162)
North America	29	1.074 (0.937-1.232)	1.004 (0.873-1.154)	0.898 (0.775-1.039)
Europe and New Zealand	20	1.174 (1.007-1.369)	1.092 (0.934-1.277)	1.062 (0.899-1.254)
China (including Hong Kong and study LIM)	27	1.321 (1.144-1.524)	1.239 (1.071-1.433)	1.175 (1.005-1.374)
Rest of Asia (including study ILCCO)	17	1.284 (1.187-1.389)	1.194 (1.103-1.291)	1.164 (1.072-1.262)
North America, Europe and New Zealand	49	1.112 (1.004-1.231)	1.037 (0.935-1.150)	0.959 (0.858-1.072)
Asia	44	1.314 (1.199-1.439)	1.229 (1.121-1.348)	1.181 (1.070-1.304)
Published in 1980s	26	1.361 (1.216-1.522)	1.267 (1.132-1.417)	1.194 (1.059-1.347)
Published in 1990s	27	1.152 (1.016-1.305)	1.077 (0.948-1.225)	1.005 (0.871-1.160)
Published in 2000s	26	1.240 (1.105-1.392)	1.163 (1.034-1.308)	1.115 (0.987-1.260)
Published in 2010s	14	1.139 (0.945-1.372)	1.059 (0.877-1.277)	1.026 (0.844-1.247)
< 100 cases	49	1.339 (1.178-1.521)	1.249 (1.098-1.422)	1.192 (1.038-1.370)
100-199 cases	22	1.117 (0.973-1.284)	1.042 (0.904-1.200)	0.978 (0.846-1.131)
200-399 cases	13	1.363 (1.190-1.561)	1.275 (1.114-1.460)	1.226 (1.051-1.429)
400+ cases	9	1.101 (0.973-1.247)	1.027 (0.905-1.166)	0.957 (0.826-1.108)
With dose-response data4	24	1.308 (1.181-1.449)	1.226 (1.105-1.359)	1.170 (1.052-1.302)
Without dose-response data	69	1.182 (1.088-1.286)	1.104 (1.014-1.201)	1.040 (0.948-1.141)
With age adjustment5	75	1.184 (1.100-1.274)	1.106 (1.027-1.191)	1.048 (0.966-1.136)
Without age adjustment	18	1.437 (1.194-1.728)	1.340 (1.110-1.618)	1.264 (1.026-1.556)
Case-control studies	77	1.226 (1.133-1.326)	1.144 (1.057-1.239)	1.080 (0.990-1.177)
Prospective studies	16	1.187 (1.043-1.350)	1.111 (0.977-1.264)	1.064 (0.928-1.220)

¹All analyses use random-effects models;

²Adjusted for confounding by fruit, vegetables and dietary fat consumption and by education;

³Using the Lee and Forey method[22] with an additive model and assuming a concordance ratio of 3 and misclassification rates of 2.5% for studies in North America and Europe and 10% for studies in Asia;

⁴Specifically for smoking by the husband;

⁵Or matching (within nonsmokers).

Table 7 Adjusted/corrected analyses: Per 10 cigs smoked by husband¹.

Studies	n	Unadjusted for confoundingUncorrected for misclassificationRR (95%CI)	Adjusted for confounding2Uncorrected for misclassificationRR (95%CI)	Adjusted for confounding2Corrected for misclassification3RR (95%CI)
All	93	1.102 (1.065-1.140)	1.062 (1.027-1.099)	1.032 (0.994-1.071)
North America	29	1.037 (0.977-1.101)	1.006 (0.946-1.070)	0.957 (0.896-1.022)
Europe and New Zealand	20	1.060 (0.995-1.128)	1.020 (0.956-1.088)	1.003 (0.938-1.073)
China (including Hong Kong and study LIM)	27	1.169 (1.082-1.263)	1.127 (1.041-1.219)	1.094 (1.006-1.191)
Rest of Asia (including study ILCCO)	17	1.142 (1.095-1.191)	1.094 (1.050-1.141)	1.079 (1.033-1.127)
North America, Europe and New Zealand	49	1.046 (1.001-1.094)	1.012 (0.967-1.059)	0.974 (0.928-1.023)
Asia	44	1.158 (1.104-1.216)	1.113 (1.060-1.170)	1.089 (1.033-1.147)
Published in 1980s	26	1.148 (1.092-1.207)	1.105 (1.052-1.162)	1.075 (1.019-1.134)
Published in 1990s	27	1.063 (1.004-1.125)	1.025 (0.967-1.087)	0.988 (0.926-1.053)
Published in 2000s	26	1.123 (1.056-1.194)	1.085 (1.020-1.155)	1.061 (0.995-1.132)
Published in 2010s	14	1.073 (0.970-1.188)	1.032 (0.932-1.143)	1.014 (0.912-1.128)
< 100 cases	49	1.143 (1.077-1.213)	1.101 (1.036-1.169)	1.072 (1.005-1.144)
100-199 cases	22	1.062 (0.993-1.137)	1.025 (0.957-1.098)	0.994 (0.926-1.066)
200-399 cases	13	1.176 (1.097-1.261)	1.134 (1.058-1.216)	1.111 (1.027-1.202)
400+ cases	9	1.041 (0.976-1.111)	1.002 (0.938-1.070)	0.966 (0.895-1.042)
With dose-response data4	24	1.123 (1.072-1.176)	1.082 (1.032-1.134)	1.053 (1.005-1.103)
Without dose-response data	69	1.091 (1.044-1.139)	1.053 (1.007-1.100)	1.021 (0.973-1.071)
With age adjustment5	75	1.084 (1.046-1.123)	1.044 (1.008-1.082)	1.015 (0.976-1.056)
Without age adjustment	18	1.211 (1.101-1.331)	1.168 (1.061-1.285)	1.131 (1.018-1.256)
Case-control studies	77	1.106 (1.064-1.150)	1.066 (1.025-1.109)	1.034 (0.991-1.080)
Prospective studies	16	1.081 (1.021-1.145)	1.043 (0.985-1.105)	1.018 (0.957-1.083)

¹All analyses use random-effects models;

²Adjusted for confounding by fruit, vegetables and dietary fat consumption and by education;

³Using the Lee and Forey method[22] with an additive model and assuming a concordance ratio of 3 and misclassification rates of 2.5% for studies in North America and Europe and 10% for studies in Asia;

⁴Specifically for smoking by the husband;

⁵Or matching (within nonsmokers).

As shown in Table 6, adjustment for confounding variables reduces the overall RR for smoking by the husband from 1.219 (1.138 to 1.305) to 1.139 (1.062-1.221), implying bias due to failure to control for the four variables is 1.219/1.139 = 1.070. Further correction for misclassification reduced the estimate to a marginally nonsignificant 1.077 (0.999-1.162), implying a further bias of 1.139/1.077 = 1.058. In the fully adjusted and corrected analyses, there is no evidence of an association in North America, Europe and New Zealand (RR 0.959, 0.858-1.072) but there is an association in Asia (1.181, 1.070-1.304).

RRs are higher for studies providing dose-response data (1.170, 1.052-1.302) than for other studies (1.040, 0.948-1.141), and higher for studies which did not adjust for age (1.264, 1.026-1.556) than for those which did (1.048, 0.966-1.136). However, neither difference is statistically significant (P = 0.10 and P = 0.08 respectively).

The pattern of results shown in Table 7, where RRs are per amount smoked by the husband, is similar, though the RRs themselves are lower. Thus, the unadjusted/uncorrected overall RR of 1.102 (1.065-1.140) reduces to 1.062 (1.027-1.099) after adjustment for confounding (bias = 1.038), and to a nonsignificant 1.032 (0.994-1.071) after further correction for misclassification (additional bias = 1.030). Patterns of variation by study factors are very similar to those for overall smoking by the husband in Table 6.

Additional material presented in http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%202.pdf shows that the effect of confounder adjustment was greatest for education, and least for fruit and vegetables. Thus, in the analysis of RRs per amount smoked by the husband, the biases due to uncontrolled confounding were estimated as 1.024 for education, 1.013 for dietary fat, 1.005 for fruit, and 1.004 for vegetables.

DISCUSSION

Introduction

We have demonstrated, as other reviews before us[6,76], a weak but significant (P < 0.05) association of ETS exposure with never smoker lung cancer risk. This can be seen for various indices of exposure, including spousal, household, workplace and total exposure. It is less clearly evident for exposure in travel and in social situations, where data are quite limited, and for childhood exposure where the results shown in Table 4 are rather conflicting. There is also clear heterogeneity between study-specific estimates for many of the indices of exposure. Meta-analyses for smoking by the spouse (or nearest equivalent) shown in Table 3 indicate that estimates are higher in early studies (published in 1981-89), in small studies (1-49 cases), and where estimates are not age-adjusted.

Do these quite weak associations provide good evidence of a causal relationship? To gain insight into this we carried out additional analyses for smoking by the husband investigating biases due to uncontrolled confounding by education and three aspects of diet (fruit, vegetables and dietary fat) and due to failure to adjust for misclassification of smoking by the subject. Based on 93 studies, confounder adjustment and misclassification correction substantially reduced the magnitude of the association with smoking by the husband, the RR (95%CI) estimate of 1.22 (1.14-1.31) reducing to 1.14 (1.06-1.22) after confounder adjustment, and further reducing to 1.08 (0.999-1.16) after additional correction for misclassification. The adjusted and corrected estimate is not quite significant, the same being true for analyses based on the RR per 10 cigs/day smoked by the spouse, where the overall RR reduced from 1.10 (1.07-1.14) to 1.06 (1.03-1.10) after adjustment for confounding and to 1.03 (0.994-1.07) after the further correction for misclassification.

Below we discuss some aspects of the evidence relevant to consideration of causality. Parts of the discussion are quite brief, the interested reader being referred to our publication[2] describing our earlier analyses.

Plausibility

Since active smoking causes lung cancer, and since ETS contains many of the carcinogens in tobacco smoke, it can be argued that some causal effect of ETS exposure is to be expected, though this argument depends on there being no threshold dose of exposure. If there is no threshold, what effect might one expect? Certainly, exposure from ETS is much less than that from active smoking, with studies based on cotinine indicating relative exposure factors of 0.4%[77], 0.2%[78] or 0.06%[79] and studies based on particulate matter[80-88] suggesting a lower factor, of order 0.005%-0.02%. Given an RR for current vs never smoking of 8.43, as reported in a recent meta-analysis[89] and assuming a linear dose-response relationship, even a relative exposure factor as high as 0.5% would only suggest that the RR for ETS exposure would be about 1.04, while a relative exposure factor of 0.1% would suggest a RR of about 1.008. These RRs are much less than the unadjusted/uncorrected RR of 1.22 for smoking by the husband (or nearest equivalent) shown in Table 6. Either the relationship between dose and risk is distinctly non-linear (and the evidence does not suggest this for active smoking[89]) or a substantial part, if not all, of the observed association is due to bias.

Confounding

Based on the evidence we collected, we have demonstrated a clear tendency for increased dietary fat consumption, reduced fruit and vegetable consumption and fewer years of education to be associated both with increased lung cancer risk and with increased at home ETS exposure. Given that relatively few of the studies adjusted for the dietary variables or education, it was to be expected that adjustment for these four factors would reduce the RR for smoking by the husband, and so it proved. While there is uncertainty in this adjustment, as discussed elsewhere[19], it is clear that there is a considerable potential for bias. Among other things it should be noted that these are not the only potential sources of bias. We considered various other candidate confounding factors, including income, occupation, and socioeconomic factors, obesity, physical activity, air pollution, alcohol and tea drinking, but concluded that for none of these were there data adequate to provide any sort of reliable qualitative estimate of their relationship to lung cancer risk in non-smokers. That said, the general tendency for smoking and marriage to a smoker to be associated with lifestyle factors generally considered associated with adverse health[90-93], suggests that our adjustments may well have been conservative.

Misclassification of active smoking

Some current or former smokers are known to deny having smoked, so being wrongly described as never smokers[26,94]. Also, marital partners’ smoking habits are correlated, with smokers tending to marry smokers[3,23]. Taken together, these two tendencies, if ignored, will bias the observed association of smoking by the husband to never smoker lung cancer risk[3,21,95]. There are many difficulties in accurately estimating the extent of bias due to misclassification. These include the misclassification rates being dependent on the circumstances in which the questions were asked, as well as the fact that smokers who deny smoking are unrepresentative of all smokers, tending to be more often occasional smokers and long-term ex-smokers and so have lower lung cancer risks than non-misclassified smokers[23]. Here we have assumed, as earlier[22], that misclassification correction can be carried out assuming that, among women, the percentage of average-risk ever-smokers who deny smoking is 10.0% in Asia and 2.5% elsewhere, these misclassification rates taking account of the lower lung cancer rates in misclassified compared to non-misclassified smokers.

While the misclassification correction is clearly open to question, and we have not formally updated the extensive work we did some years ago on estimating rates[23,26], we still believe that the rates we have used are not unreasonable. Indeed given recent estimates of substantial denial of smoking in recent studies[94,96,97], our correction may be somewhat conservative.

We now briefly comment on other sources of bias.

Publication bias

Publication bias occurs if the data that are published are not representative of all the data that exist on a topic. For many exposures, positive findings are published more often than negative findings[98-100], so meta-analyses of data drawn from the literature overestimate true relationships. We have not attempted to quantify the extent of publication bias, though our detailed tables (http://www.pnlee.co.uk/downloads/etslc/23482-supplementary%20file%204.pdf) do include results of Egger tests[16], a number showing some evidence that smaller studies are more likely to produce above average than below average RRs. This is consistent with the higher RRs reported in small studies seen in Table 3 for spousal smoking. We believe that some publication bias exists but given that the larger studies seem likely to publish regardless of the findings, and that these contribute most to the overall estimates, such bias is probably unimportant.

There is some evidence (P = 0.10) that RRs are higher for those studies which provide dose-response results than for those which do not so. If this represents a true effect, it is suggestive of a different form of publication bias, with authors tending to be more likely to report dose-response results where there is a strong association in the first place.

Diagnostic inaccuracy

Misdiagnosis of lung cancer certainly exists, especially when based on X-rays or sputum cytology[101-103]. The extent, and direction, to which it might have biased the RR estimate for ETS and lung cancer is difficult to determine. While randomly misdiagnosing as lung cancer diseases which are unassociated with ETS would tend to dilute any true RR, misdiagnosis might not be random and may be correlated with ETS exposure or factors associated with it. Since random-effects estimates for spousal smoking proved to be quite similar for studies that did or did not require full histological confirmation, this seems unlikely to be an important source of bias.

Errors in determining ETS exposure

Case-control studies collect exposure data after the disease has occurred, and the presence of the disease itself, or knowledge of it, may distort responses about past exposure. Such recall bias is not an issue for prospective studies. Given that our analyses for spousal smoking found little difference in RRs by study type, we feel that recall bias is unlikely to be a major problem.

Random misclassification of smoking spouses as non-smokers will not create a false effect if no true risk exists, but will underestimate a true relationship. It has been clearly shown[21] that such misclassification causes much less bias effect than does misclassification of the subject’s smoking, so for practical purposes it can be ignored.

Bias from ETS exposure in the reference group

When considering the relationship of lung cancer risk to smoking by the husband, three categories of women are relevant: A - never smokers married to ever smokers; B - never smokers married to never smokers; and C - never smokers without any ETS exposure. Group C is a subset of group B. In the analysis of the effect of husband’s smoking, the RR estimate is based on comparison of groups A and B, but it has been argued[3] that a better estimate RR* is based on comparison of groups A and C. If a marker of ETS exposure, such as cotinine, is Z times higher in group A than group B, RR* can be estimated by RR* = RR(Z-1) / (Z-RR)[2,3].

Some comments can be made on this revised estimate. First, and most noteworthy, to conduct background correction only makes sense when the original association, with marriage to a smoker, derives from a causal effect of ETS. Where adjustment for confounding and correction for smoker misclassification bias explains the whole of the observed association, background correction will have no effect. If such adjustment and correction explains most of the association, the correction will have a small effect. Thus, assuming Z = 3, as estimated by Hackshaw et al[3], this correction has quite a substantial effect on the unadjusted association for husband’s smoking, increasing it from 1.22 to 1.37. However, applying it to the confounder adjusted and misclassification corrected estimate of 1.08 only increases it to 1.12. In any case, the validity of the background-corrected estimate of 1.12 is dubious, given that the 1.08 was not statistically significant in the first place, and could itself be an overestimate due to the limitations in confounder adjustment and misclassification correction discussed above.

Second, background correction only applies to the simple comparison of risk in the exposed and comparison groups, and does not apply to estimates of the increase in risk for amount smoked by the husband. Also, background correction is only an indirect method for estimating lung cancer risk from sources of ETS exposure other than the spouse, using data only relating to spousal exposure. This method ignores existing data on risk from these other sources.

Overall impression

Coming to reliable conclusions regarding a weak association based on non-randomized epidemiological studies is difficult at the best of times. When, as here, there is evidence that adjustment for confounding and correction for misclassification substantially weakens the association most usually considered (smoking by the husband) and renders it nonsignificant, and when these adjustments and corrections may themselves be somewhat limited, one cannot reliably conclude that a true effect of ETS exposure on lung cancer risk has been demonstrated. While one cannot prove a negative, and while the clear relationship of smoking to lung cancer suggests that some association might exist, the only conclusion that seems valid is that there may be a relationship of ETS to lung cancer risk (with the evidence stronger for Asian studies), but if it exists, it is certainly much weaker than suggested by meta-analyses that do not adjust for confounding and misclassification.

Most, if not all, of the weak association of ETS with risk of lung cancer is explicable by confounding and smoking misclassification. A causal relationship is not demonstrated.

ACKNOWLEDGMENTS

We thank Japan Tobacco International S.A. for supporting publication of this paper. The opinions and conclusions of the authors are their own, and do not necessarily reflect the position of Japan Tobacco International S.A. We also thank the United Kingdom Tobacco Manufacturers Association, Imperial Tobacco Ltd, British-American Tobacco Limited, and Philip Morris Products S.A. for earlier support in developing the databases used. Finally we also thank Pauline Wassell, Diana Morris and Yvonne Cooper for assistance in typing various drafts of the paper and obtaining relevant literature, and all the researchers who published the reports which formed the basis of our work.

COMMENTS

Background

The authors address the widely held claim that environmental tobacco smoke (ETS) exposure causes lung cancer by presenting an up-to-date comprehensive meta-analysis of the available evidence, considering exposure from various sources, and illustrating the potential magnitude of bias from confounding and misclassification of smoking.

Research frontiers

Based on all 102 studies providing relevant data, the authors demonstrate a significant (P < 0.05) increase in never smoker lung cancer risk for various exposure indices - from the spouse, at home, at work and overall, though the evidence for childhood exposure is less clear. Based on smoking by the husband, the most studied ETS exposure index, the RR is estimated as 1.22 (95%CI: 1.14-1.31). However, adjustment for confounding by education and by consumption of fruit, vegetables and dietary fat, and correction for misclassification of active smoking by the wife, markedly reduces this association, which becomes a nonsignificant 1.08 (95%CI: 0.999-1.16). Since these adjustments and corrections may not fully correct for the bias from these sources, and given the existence of other biases, one cannot conclude with any certainty that a true effect of ETS exposure on lung cancer risk exists.

Innovations and breakthroughs

The new feature of the paper is the extent of the evidence considered, and the adjustments and corrections made.

Applications

The authors analysis should engender caution in drawing inferences from weak associations seen in non-randomized epidemiological studies, particularly where biases are known to exist.

Peer-review

It is an interesting paper.