Tobacco Documents Online

BEOCLaTOaY TOXICOLOGY dS0 PHAft}fACOLOGY 2I, 18-1-191 (19951 Publication Bias in the Environmental Tobacco Smoke/ Coronary Heart Disease Epidemiologic Literature' MAURICE E. LEVOfS' AVD MAXWELL W. LAYARDt 'Eneironmenta/ Health Resources, Tiburon. California 94920: and }Ipyard Associates. Alamrda. Catifornia 94501 Received June 11, t994 the environmental tobacco smoke/coronary heart dis- ease (ETS/CHD) literature: (1) Statistical tests applied to all sez-specific relative risk (rr) estimates from 14 previously published studies indicate that publication bias is likely. A funnel graph of the studies' log relative risks plotted against their standard errorsis asymmet- rical, and weighted regression of the studies' log rela- tive risks on their standard errors issigni8caat (P < 0.01). (2) Previously unpublished ETS/CHD relative risks from the American Cancer Society's Cancer Pre- vention Studies (CPS-I and CPS-U) and the National Mortality Followback Survey (NMFS) do not show an increased CHD risk associated with ETS exposure. CPS-I: men, rr = 0.97 (0.90-1.05); CPS-I: women, rr = 1.03 (0.98-1.08); CPS-II: men, rr = 0.97 (0.87-1.08); CPS-II: women, rr = 1.00, (0.88-1.14); NMFS: men, rr = 0.97 (0.73-1.28); women, rr = 0.99 (0.84-1.16). Comparison of pooled relative risk estimates from 14 previously published studies (rr = 1.29; 1.18-1.41) and unpublished results from three studies (rr = 1.00; 0.97- 1.04) also indicates that published data overestimate the association of spousal smoking and CHD (Xt = 25.1; P< 0.0001). t te9a Ar.dede Pr<ee. t.e. - Publication bias is the systematic error in the pub- Two approaches are used to assess publication bias in lished literature produced when the results of studies in- INTRODUCTION I 1 Many papers have appeared in recent years address- ing the problem of publication bias (Rosenthal, 1979; Simes, 1986a,b; Chalmers et a1, 1987, 1990; Dickersin, 1990; Dickersin et aL, 1987, 1992; Liglit, 1987; Begg and Berlin, 1988; Peto, 1992). Nearly everyone agrees that publication bias tends to distort estimates of association obtained by pooling the results of published studies (i.e., quantitative meta-analysis), so that inferences about the presence and size of associations are rarely appropri- ately conservative. It is all but unanimously agreed that publication bias is a serious problem. 'This work was supported in part by funding from Phillip Morris U.S.A. The views expressed represent the personal opinions of the au. thors and are not necessarily those of Phillip Morris C.S.A. fluence the decisions, by authors or by editors, to pub- lish. It has long been suspected that chance, together with a preference for statistical significance when pub- lishing small studies, plays a major role in publication bias (Rosenthal, 1979). However, the publication pro- cess is complex and is affected by the preferences of funding agencies, editors, and authors. Thus, publica- tion of industry-financed research on environmental exposures may be conditioned on finding null results (Brosa,1981; Kotelchuk,197d), publication of drug com- pany-financed research may be conditioned on finding benefit from a new therapy (Davidson, 1986), and publi- cation of agency-financed research may be conditioned on support for the agency's goals and objectives (Rennie and Flanagin, 1992). In addition, bias in favor of politi- cally correct results may be a growing problem (Smith, 1980; Glass et al., 1981; Feinstein, 1992). Editors once openly stated a preference for statisti- cally significant results (Melton, 1962). Today it is rec- ognized that such standards are certain to produce pub- lication bias, and editors are more inclined to publish null studies than they were a few years ago, especially if they contradict an earlier study and are of "equal or superior quality" (Angell, 1989). However, studies re- porting null results will always have to meet a higher editorial standard of excellence than positive studies be- cause, at a minimum, null results require high statistical power in order to be interpreted at all. It is probably safe to say that no scientific publication is free ofbias, There- fore, we must learn to identify the effects of publication bias and adjust our inferences accordingly. There are both empirical and methodological grounds for suspecting that publication bias may have inflated relative risk estimates derived from meta-analysis on re- sults of published epidemiologic studies on ETS and var- ious disease endpoints (Vandenbroucke, 1988; Peto, 1992; Dickersin and Berlin, 1992). In the case of coro- nary heart disease (CHD) most of the ETS/CHD studies are,quite small, as seen in Table 1. Begg and Berlin(198$) noted that the presence of several small studies I 0"3.2300/95 $6.00 194 Copyright { 1995 by Academic Press. Inc. All rights of reproduction in any form reserved.

PUBLICATION BIAS IN ETS HE.4RT. DISEASE EPIDEMIOLOGY TABLE 1 Previously Published ETS-Heart Disease Epidemiologic Studies 185 95 ~ Cl Study Sex' No. cases . . Z'n:Ex° . ETS bv design3 Relative risk Lower - C pper 1. Butler(1988) . F - 60:4 -- Yes - i.40 0.51 3.84 2. Dobsoneral.t199t1 F tG:43 Yes 2.46 1.47, 4.13 M 161:'?? 0.9i 0.50 1,86 3. Garland et aL (198.5) F . 2:1; No 2.70 0.90 13 60 4. He et al. 11989) F 9r?5 . I ? 1.50 0.9 . 2.51 5. Heeta/.(1994) F 11:15 Yes 1.24 0.56 2.72 6. HelsingeOal.(1988) F 43;:551 No 1.24 11 1,4 M 248:1':2 1.31 1.1 L6 Sandler and Shore /1989) F (Satne) :vo 1.19 1.04 1.36 M ~-~ - 1.31 1.05 1,64 7. Hirayama(1984) F 1I8:376 No . . . 1.15 0.94 1.42 8. Hole et al. (1989/ F& M 84 No . 2.01 1.21 J.35 Gillisetaf.(1984) F 2:19 - 3.56 0.83 15.4 M 18:14 . . 1.29 0.64 2.64 9. Humblectat.(1990) F 27:49 No 1.59 0.99 2.5; 10. Jackson (1989) F 20 Yes 4.00 1.35 13.1 M 49 1.10 0.40 3.00 il. LaVecchiaetaf.(1993) F 44 .- 6 Yes 1.19 0.49 :.8' . M 9 .. 1.43 . 0.59 2.94 12. 1" et o1. (1986) F 22:55 ~ Yes 0.97 0.56 1.69 M 26:15 1.34 0.64 ?.80 13. Siartin et at. (1986) F 23 2.6 L'20 u.CO 14. Svendseneta(. (1987) M 8:5 . . No 2.^_3 0.12 6.92 Summaryresulu` - 1.29 1.18 1.41 ° Number of hearr disease cases-ETS unexposed:ETS exposed. • Original design of study. . ` Most recent results are used in cases of multiple reporting. on the same issue increases the risk of publication bias. Lee (1992) observed that the ETS/CHD literature shows signs of publication bias, noting that in several large cohort studies there are relevant unpublished data. Bero et al. (1994) disputed this view, asserting that pub- lication bias has not influenced the ETS epidemiologic literature. The study reported here was undertaken to evaluate the ETS/CHD literature for evidence of publi- cation bias. Statistical Tests o(Publication Bias Hedges (1984), Hedges and 0lkin (1985), and Berlin et al. (1989) proposed methods of evaluating publication bias based upon truncated sampling models. lyengar and Greenhouse (1988) proposed methods based upon weighted distribution theory. These methods are com- putationally difficult, require an understanding of un- derlying models unfamiliar to many epidemiologists, and have not received much attention outside of statis- tical journals. Light and Pillemer (1984) recommended a simple and intuitively 'obvious method of evaluating publication Several statistical methods of detecting and quantify- bias that involves visual inspection for departure from ing publication bias have been recommended. Rosenthal what is termed a "funnel graph" appearance in a plot of (1979) proposed correcting pooled P values based upon the estimated size of effect against study sample size, for an estimate of the number of unpublished studiee. That all studies providing data. More recently ' Vanden- ' estimate is computed by adding the standard normalde- broucke (1988) and Berlin et al. (1989) recommended a viates associated with the Pvalues obtained and dividing refinement of the funnel graph approach that uses the by the square root of the number of studies being com- log relative risk for the estimated size of effect and the bined. This method is computationally simple, but it has standard error of the log relative risk in place of the been criticized on grounds that it forces the assumption study sample size. The two methods of constructing a of zero effect in the unpublished studies, it ignores pos- funnel graph are very similar. The latter approach has ~ sible variation in unpublished study size, and it produces neither an estimate of treatment effect nor a test of sig- nificance of the effect of publication bias (Begg and Ber- lin, 1988). the advantage of using a more appropriate measure of~ precision of the effect estimate than study sample size. ~ If a collection of studies provides an unbiased estimate of the treatment effect, then random sampling errot~~ ,ej CJ ~ ~

1S6 LEVOtS AND LAYARD 1 I I I I I I should result in an approximately symmetrical distribu- tion of results both above and below the summary rela- tive risk from the pooled data, with scatte r being greatest for the smallest studies and narrowing as the size of the studies increases. If substantial publication bias is pies- ent, then the plot will be markedly asymmetrical as study size decreases, and the top half of the funnel will contain most of the results. Each recommended method for evaluating publica- tion bias is based upon the same underlying assumption. "Our assumption is that among all studies, the effect sizes and the sample sizes should be independent ... (Berlin et al., 1989, p. 383). This assumption constitutes a null hypothesis that was tested directly in the present study by weighted regression of the ETS/CHD studies' log relative risks against the standard errors of the log relative risks. This test is computationa)ly simple, and the resulting regression coefficient provides a test of the statistical significance of publication bias. Comparison with Unpublished Data Several authors have also suggested that publication bias can be assessed by comparison of published results with results obtained from previously unpublished stud- ies (Simes, 1986a,b; Begg and Berlin, 1988; Chalmers et at, 1987; Dickersin, 1990). To make such a comparison, we analyzed data from two American Cancer Society (ACS) cohort studies, Cancer Prevention Study-I (CPS- I, sometimes referred to as the "million person study") and Cancer Prevention Study-II (CPS-II). Also, a rela- tive risk from an analysis (Layard, 1995) of the National Mortality Followback Survey (NMFS) was used. MATERIALS AND METHODS Funnel Graph Test In Fig. 1 sex-specific relative risks from all currently available ETS/CHD studies are used to produce a funnel graph by plotting log relative risks against their esti- mated standard errors (Vandenbroucke, 1988; Berlin et al., 1989). Dashed lines are used to outline an imaginary funnel that illustrates the symmetry and dispersion of results- expected from an unbiased sample of studies of different sizes. Unpublished Data (1972). Causes of death were coded with the Interna- tional Classification of Disease (ICD) Revision 7. In CPS-II approximately 1.2 million subjects (509,000 men and 677,000 women) were enrolled by ACS volun- teers in all 50 states, the District of Columbia, and Puerto Rico, in late 1982. Again, information on study factors collected by questionnaire at the beginning of the study forms the basis for the present analysis. The co- hort was followed for 6 years (1983-1988). Vital status was ascertained for 98.2% of the cohort, and death cer- tificates were obtained for 94% of decedents. Causes of death were coded with ICD Revision 9. In both CPS studies, subjects were recruited by ACS volunteers from among friends, relatives, neighbors, and other acquaintances. Although volunteers came from all social classes, and tended to recruit subjects in the same socioeconomic class as themselves, the cohorts are not completely representative of the national population. For example, representative numbers of illiterate peo- ple, institutionalized people, itinerant workers, illegal aliens, military personnel, construction workers, and people who tend to move often were not included. tiZi- nority races and inner city residents were also underrep- , resented. Because of these differences, the socioeco- nomic level of the cohort was somewhat higher than that for the nation as a whole. In addition, sick people were likely to have been underrepresented in the study sam- ples. For the reasons stated above there is an apparent healthy person effect on the CPS death rates, as the age- specific death rates are lower for the CPS cohorts than those for the national population. Although death rates are not representative of the entire U.S. population, rel- ative risks are based upon internal comparisons and should be reasonably reliable- Materials and methods employed in the conduct of the ACS studies are discussed in greater detail elsewhere (Hammond, 1966; Garfinkel, 1980; Garfinkel and Stell- man,1988). The spousal smoking definition of ETS exposure and coding of CHD mortality employed in the present anal- ysis are similar to the operational definitions of these variables used in other published ETS/CHD epidemio- logic studies. These vary somewhat between CPS-I and CPS-lI and are described in greater-detail below. Only self-reported never-smokers who had spouses with known smoking habits were used in the analyses. Both CPS-I and CPS-II collected data on self-re- i ported smoking habits in terms of cigarettes smoked per In CPS-I more than one million men and women day, which we grouped into the following categories: (456,000 men and 595,000 women) were enrolled by ACS "Ex" (former smoker), 1-19,20-39, and 40 or more. Rel- volunteers in 25 states in 1959-1960. Information...on ..ative risks for increasing spous.f smoking levels were ~ study factors collected by questionnaire at the beginning of the study forms the basis for the present analysis. The cohort was followed for 13 years (1960-1972). Follow-up was complete for 98.4% of the cohort through June 1971 and was 92.7% complete for the 13th year of the study tested for trend in both the sex- and study-specific re- 0 sulta and in the pooled results. Data on cigar and pipe ~ smoking only were also collected for men. The category ~ °Any" ETS exposure is the global spousal smoking exposure definition. It is the exposure measure reported ~

PUBLICATION BIAS IN ETS HEART DISEASE EPIDEMIOLOGY Log Relative Risk 2.0 I I r I D I I I I I , ' I I 1.8 1 1.2 0.8 0.4 0.25 --t 0.0 4.4 -1.2 • )uEn,lya fiVelef -haly ee: tillUhee A UnuublbMd t[WM• , _0.75 a/. t•5e 0.1 0.2 0.3 0.4 0.5 0.6 Standard Error of Log Relative Risk 0.7 187 FiO. 1. Funnel graph. Sex•specific log relative risks for each study are plotted against their standard erron. Unpublished results are depicted as triangles. Previously published results are shown as squares. The solid horizontal line shows the summary log relative risk for the pooled published studies (rr = 1.29: log rr = 0.25). Dashed lines illustrate the expecteddispersion of study results in the absence of bias. in the largest number of published studies and is the definition used to summarize the sex- and study-specific relative risks and for pooling the unpublished results. International Classification of Disease codes were used by ACS to code the cause of death listed on death certificates of deceased cohort members in both studies. These codes underwent two revisions between CPS-I and CPS-II. The definition of heart disease death for the CPS-I analysis includes the following ICD-7 codes: 420.0-420.2. The definition of heart disease death for the CPS-II analysis includes the following ICD-9 codes: 410-414. Both ICD code ranges primarily cover arterio- sclerotic heart disease,. myocardial infarction, and an- gina pectoris. Layard (1995) reported results from an analysis of data from the NMFS. These data were collected by the National Center for Health Statistics in 1986. The NMFS.was a representative 1% sample of U.S. adult deaths (>25 years). The 1986 Current Mortality Sample, a systematic sample of death certificates sent by state vital statistics offices to NCHS approximately 3 months after death, was used to select the sample for the NMFS. Materials and methods employed in the conduct of the NMFS are discussed in greater detail elsewhere (See- man et al., 1989). In the CPS-I and CPS=II analyses the relative risk of death from CHD among never-smokers married to smokers compared to never-smokers married to never- smokers was calculated for men and women separately using Poisson regression methods (Breslow and Day, 1987). All CPS-1 and CPS-II results were stratified by sex and adjusted for age and race. Relative risks were combined by computing a weighted average of the log relative risks, the weights being the inverse of the log relative risks (Woolf, 1955). The difference between the pooled relative risks for published and unpublished results was tested for statis- tical significance by means of a X2 test. RESULTS FunnelGroph Figure 1 presents the sex-specific log relative risk for each study plotted against its standard error. Unpub- lished results are depicted as triangles. Previously pub- lished results are shown as squares. The solid horizontal line shows the summary log relative risk for the pooled published studies (.rr. = 1.29; log n= 0.25). Dashed lines illustrate the expected dispersion of study results in the absence of bias. The plot shown in Fig. I is not symmetrical. Most of the results to the right of the unpublished results are in the top half of the funnel. As the standard error of the log relative risk becomes large, all of the published re- sults are above the summary log reteiiive risk. Regression Analysis Weighted regression of the log relative risks on their standard errors is statistically significant (P <0.01).

188 LEVOIS AND LAYARD TABLE 2 Age and Race of Never-Smoking Men and Women and the Smoking Status of Their Spouses in CPS-I and CPS-li Never-smoking men " Never-smkingwomen CPS-1 I I I I I I Mean age at entry 56.1 years Mean age at entry 52.8 yean Race 85.811 (97.0%) White' Race 259,429497.0%) White 1,704 (1.9%) Black 4.852(1.8tlF) Black 943 (1.1°b) Other 3,131 (1.2%) Other Total 88,458 (100.0%) Total 267,412 (100A%<) Smoking status of wife 73.890(83.5%) Never Smoking status of husband 73,895 (27.8%) Never 14,568 (16.5%) Ever 193,517 (72.4 S o) Ever Total 88.458 (100.04'a ) Total 267,412 (100:0%) Mean age at entry CPSd1 57.7 years Mean age at entry 55.8 years Race 98,579 (95.0%n) White Race 215,132/95.2'8) White 2,745 (2.6`a) Black 6,018/2.6°.b) Black . .. 2,448 42.4%) Other 4,917/2.2:G) Other Total 103,772 (100.0%) Toesf 226,067 (100.0%) Smoking status of wife 77,339 (74.5%) Never Smoking statue of husband 77,455 (34.3%) Never 28,433 (25.5%) Ever 148,612 (65.7 n) Ever Total 103,772 (100.0%) Total 226,067 (100.0%) Comparison with Unpublished Results Table 2 gives the mean age and race of never-smoking men and women, and the smoking status of their spouses, in CPS-1 and CPS-II. Table 3 gives the number of CHD deaths among never-smoking men and women in the two cohorts grouped according to the smoking sta- tus of the spouse. In the CPS-I cohort there was a total of 88,458 male and 267,412 female never-smokers with spouses having known smoking habits. Among these subjects, there were 7758 CHD deaths in males and 7133 CHD deaths in females. In the CPS-II cohort there was a total of 103,772 male and 226,067 female never-smokers with spouses having known smoking habits. Among these subjects there were 1966 CHD deaths in males and 1099 CHD deaths in fe- males. Table 4 presents the relative risks and 95%a confidence intervals calculated from the CPS study data. Relative risks were adjusted for age and race. Further adjustment using a weight index, exercise, highest level of education, dietary factors, alcohol consumption, history of hyper- tension, and history of diabetes had no appreciable effect on any of the reported associations. It is evident from Table 4 that most of the relative risks are very near 1.00, regardless of sex or spousal smoking behavior. There are four relative risk estimates with confidence intervals that exclude 1.00, all of which are in men. In the CPS-II cohort, never-smoking men with exsmoking wives expe- rienced significantly lower CHD death rates than never- smoking men with never-smoking wives (rr = 0.81, CI TABLE 3 0.70-0.93). Among men with wives who smoked 1-19 Deaths from Coronary Heart Dlsease among Never- cigarettes per day, the relative risk was rr = 1.36 (1.10- SmokersGroupedbyCigaretteeperDaySmokedbythe 1.68), but lower, a= 1.26 (1.00-1.58), among men with Spouse in CPS-I and CPS-11 wives who smoked 20-40 cigarettes per day, and still - Cigaretta par day smokld lower, rr = 1.13 (0.61-2.11), among men with wives who by the apouae smoked 40+ cigarettes per day. No significant trend was None Et 1-19 20-39 40+ Pipe/ ciger` CPS-1 Men 6954 206 400 186 0 Women 2217 1685 949 980. 1 1192 CPS-fl Men 1566 223 90 77 10 0 Women 376 470 56 60 19 118 ' Quit smoking before the beginning of the study. ° Smoked pipes and cigars only. I Total 7758 7133 1966 1099 observed in any of the separate or combined results. After the relative risks for the two cohorts were com- bined, only the reduction in risk among men married to essmokers was significant (rr = 0.88, CI 0.79-0.97). No relative risk was significant for any women married to a smoker or for any level of exposure after both sexes, and both studies, were combined. Table 5 gives the number of never-smoking male and female cases and controls in the NMFS analysis accord- ing to the smoking status of their spouses, as well as the

PCBLICATION BIAS IN ETS HEART DISEASE EPIDEMIOLOGY 189 TP.BLE e viously unpublished data from three studies indicate Relative Risks for Death from Coronary Heart Dis- that publication bias is present in the ETS/CHD litera- ease According to Cigarettes per Day Smoked by the ture. A funnel graph of study results does not exhibit SpouseinCPS-IahdCPS-II' the symmetry expected from an unbiased collection of Men Women Exposure rr 95% CI Exposure rr CPS-I Ex" 0.95 0.83, 1.09 Er' 0.99 1-19 0.99 0.89, 1.09 1-19 1.04 20-39 0.98 0.85, 1.13 20-39 1.06 40+ 0.72 0.41, 1.28 40+ 0.95 Any 0.97 0.90, 1.05 P/cigar` 1.06 Any 1.03 CPS-II Er' 0.81 070, 0.93 E:' 0.99 1-19 1.36 L10, 1.68 1-19 1.14 20-39 1.26 1.00, 1.58 20-39 0.98 40+ 1.13 0.61, 2.11 40+ 1.27 Any 097 0.87,1.08 P/cigar` 0.98 Any 1.00 Combined Ex° 0.79 0.80,0.97 Ec' 0.99 1-19 20-39 1.05 1.06 0.96, 1.15 0.93, 1.19 1-19 20-39 1.05 1.06 40+ 0.89 0.58,1.35 40+ 0.99 Any 0.97 0.91, 1.03 P/cigar` 1.05 Any 1.02 Both sexes, both cohorts, combined 95eC1 studies. Study size, expressed as the standard error of the log relative risk, and estimated effect size, expressed as the log relative risk, were found to be highly signifi- cantly correlated in a weighted regression analysis (P < 0.01) of the previously reported ETS/CHD study results. 0.93,1.05 Both observations support the inference that publica- 0.97,1.12 tion of ETS/CHD results is more likely if the results are 0.98. 1.15 positivethan if they are negative or null. 0.78,1.15 Comparison of published results with previously un- 0.99, 1.14 . . 0.98,1.08 published data from three large studies provides addi- ...:...:..-..tional support for this conclusion. Meta-analysis of re- suits from 14 currently published ETS/CHD epidemio- o.e6,1.13 logic studies produced a pooled relative risk estimate of 0.86,1.51 rr = 1.29 (1-18, 1.41). Meta-analysis of previously un- 0.75,1.29 published results from the two large ACS cohort studies, 0.80, 2.01 0.79 and the NMFS, produced a pooled relative risk estimate , 1 . 20 0.88.1.14 of rr = 1.00 (0.97, 1.04). The difference between these two pooled relative risk estimates is highly significant (XL = 25.1; P< 0.0001). This discrepancy between the 0.93. i.os relative risk estimates derived from published and un- 0.97;Y.13 published data provides further support for the infer- 0.98,1.14 ence that publication of ETS/CHD results is more likely 0.83, 1.18 09g 112 if the results are positivee than if they are negative or 0.98, 1.07 . null. . Ez' 1-19 . 0.96 1.05 0.91,1.01 0.99, 1.11 20-39 1.06 . 0.99, 1.12 40+ 0.97 0.93, 1.15 Any 1.0.0 0.97,1.04 . ' Adjusted for age and race. ° Quit smoking before the beginning of the study. ` Smoked pipes and cigan only. Given the strong indications of bias in the published literature, and the complete absence of association be- tween ETS and CHD observed in previously unpub- lished results described here, it is possible that publica- tion bias alone could account for the 29% excess risk re- ported in the published literature. This conclusion is in sharp contrast to the conclusion of Bero et al. (1994) that "There is no publication bias against statistically nonsignificant results on ETS in the peer-reviewed literatui•e: " associated odds ratios and 95% confidence intervals (La- yard, 1995). There was a total of 475 CHD deaths among TABLE 5 never-smoking men in the case group and 998 nonamok- National Mortality Followback Survey Coronary ing-related deaths among never-smoking men in the Heart Disease/Environmental Tobacco Smoke Case- control group. There was a total of 914 CHD deaths Control Studya among never-smoking women in the case group and 1930 nonsmoking-related deatha among never-smoking women in the control'group. In both men and women, the risk of CHD death was nearly the same regardless of the smoking status of the spouse (men, rr = 0.97 (0.73- 1.28); women, rr = 0.99 (0:54-1.16)). The pooled relative risk for published results, n=1.29 (1.18-1.41), is significantly higher than the pooled rela- tive risk for the unpublished results, rr = 1.00 (0.97- 1.04); X' = 25.1 (P < 0.0001). DISCUSSION AND CONCLUSIONS Statistical tests performed on results from 14 pub- lished ETS/CHD epidemiologic studies and on pre- Spousal smoking - Controln rt• 95~1 CI No 379 Men' 7g3 1.0 Yes 97 215 0.97 0.73-1.28 No 459 Women` 969 , 1.0 Yes 455 961 0.99 0.84-1.16 ' Layard (1995). ° Adjusted for age. `Never-amokere.

.+ 190 LEVOIS AND LAYARD Reasons for publication bias in the ETS/CHD litera- ture are unclear. Many factors could cause publication bias in this field. It is generally assumed that both au- thors and editors favor statistically significant study re- sults, particularly if the study is small, and that this pref- erence accounts for most publication bias. However, most of the ETS/CHD studies are small and report non- significant results, so achieving statistical significance alone cannot account for the observed publication bias. Another possible explanation for the observation of publication bias in the ETS/CHD literature is that there is relatively ample institutional financing of tobacco-re- lated health effects research and virtually no institu- tional support for discussing contrary findings when re- porting results. Given the large, and rapidly growing, number of studies with data that could be tested for an ETS/CHD association, and the willingness of authors and editors to publish small positive studies in this field, the effects of publication bias in the ETS/CHD litera- ture are likely to become even greater in the future. REFERENCES Angell, M. (1989). Negative studies (editorial). N Engi J. Med 321, 464-466. Begg, C. B., and Berlin, J. A. (1988). Publication bias: A problem in interpreting medical data. J. R. Stat. Soc. 151, 419-463. Berlin, J. A.. Begg, C. B., and Louis, T. A. (1989). An assessment of publication biae using a sample of published clinical trials. J. Am. Stat Assoc. 84, 38L-392. Bero, L A., Glants, S. A., and Rennie, D. (1994). Publication bias and public health policy on environmental tobacco smoke. JAMA 272, 133-136. Bmlow, N. E., and Day, N. E. (1987). The Design and Analysis of Cohort Studies. IARC, Lyon, France. Broee, 1. D. J. (1981). Scientific Strategies to Smx Your Life: A Statia- ticd Approach to Primary Preuention. Decker, New York. Butler, T. L. (1988). The Re/ationship o/Passiue Smoking to Various Health Outcomes among Seventh Day Adventists in California (Dissertation/. University of Californie, Los Angeles. Chalmers, T. C., Frank, C. S., and Reitman, D. (1990). Minimizing the three etagee of publication biea. JAMA 2 63, 1392-1395. Chalmers, T. C., Levin, H., Sacks, M. S., and Nagdingam, R. (1987). Meta-analysia of clinical triale as a scientific diecipllne•. Control of bias and comparison with large cooperative triaL. Stct. Med 8, 315- 325. Cresa, IL D., HoUy, E. A., Aaton, D. A, Ahn, K. A., and Kristainaen, J. J. (1994). Cherecterietics of women nonsmokers exposed to pae- sive smoke. Pnu. Med 23, 4U-47: Davidson, R. A. (1986). Source of Atnding and outcome of clinical tri- ala. J. Cen. Int Med. 1,155-158. Dickersin, K. (1990). The existence of QpbHcatlon bias and risk factors for its occurrence. JAMA 263, 1385-1389. Dickersin, K., Chan, S., Chalmers, T. C., Sacks. H. S., and Smith, H. (1987). Publication bin and clinical triala. Controlled Cfin.7Yiala 8, 343-353. - Dickenin, K., Yuan, -I. M., and Curtis, L. M. (1992). Factors influ- encing publication of research results. JAMA 267, 374-378. Dobson, A. J., Alexander, H. M., Heiler, R. F., and Lloyd, D. M. (1991). Passive smoking and the risk of heart attack or coronary death. Med. J. Awt. 154, 793-797, ... Feinstein, A. (1992). Critique: Juetice. science, and the bed guys. Toz. icol. PathoL 20, 303-305. Fleisa. J. L., and Grose, A. J. (1990). Meta-analysisin epidemiology, with reference to studies of the association between exposure to en- vironmental tobacco smoke and lungcancer. A ctitique. J. Clm. Epi- demiol. 44(2), 127-139. - Friedman, G. D., Petitti, D. B., and Bawol, R. D. (1986). Prevalence and correlates of passive smoking. Am. J. Public Health 73, 401- 405. Garfinkel, L. (1980). Cancer mortality in nonsmokers: Prospective study by the American Cancer Society. J. NatL Cancer fnst. 65, 1169-1173. Garfinkel. L., and Stellman, S. D. (1988). Smoking and lung cancer in women: Findings in a prospective study. Cancer Res. 48, 6951-6955. Garland, C. E., Barrett-Connoq E., Suarez, L., Criqui, M. H., and W ingard, D. L. (19g5). Effects of passive smoking on ischemic heenn disease mortality of nonsmokers: A prospective study. Am. J Epi- demiot. 121, 645-6.50. Gillis, C. R., Hole, D. J., Hawthorne, V. M., and Boyle, P. (1984). The effect ofenvironmental tobacco smoke in two urban communities in the weatof Scotlend Eur. J. Respb. Dia. 86(SuppL), 121-126. Glass, G. V., McGraw, B., and Smith, M. L. (1981). Meta-anatysis in Social Research. Sage, Beverly Hills. Hammond, E. C. (1966). Smoking in relation to the death rates of one million men and women. NatLCancer fnet. Monogr. 19, 127-204. He, Y., Lam, T. H., and Li, L. S. (1994). Passive smoking at work ae a risk factor for coronary heart disease in Chinese women who have never_smoked. Br. Med. J. 308, 380-384. He. Y., Li, C. S., Waa, Z. H., Li, S. S., Zheng, X. L., and Jia, G. L (1989). Women's paeeive smoking and coronary heart disease (English abstract only). Chung-hua Yu Fang f Hsueh Tso Chih 23, 19-22. Hedges, L. V. (1984). Estimation of effect eiu under nonrandom sam_ pling: The effects of ceneoring studies yielding statistically insig- nificant mean diHenneee. J. Educ. Stnt, 9, 61-85. Hedges, L V., and Olkin, I. (1985). Statistical Methods for Meta- . , analyair. Academic Press, Orlando. - Helsing, K., Sandler, D., Cometock, G„ and Cheq E. (1988). Heart - disease mortality in nonsmokers living with smokers. Arn J Epide- miol 127, 915-922. . H'uayama, T. (1984). Lung cancer in Japan: Effecte of nutrition and _. passive smoking. In Lury Cancer. Cautea and Preuention 1Miull and Correa, ede.). Verlag-Chemie, New York. Hole, D. J., Gillis, C. R., Chopra, C., and Hawthorne, V. M. (1989). Passive smoking and cardiorespRamry health in the weet of Scot- land. Br. Med. J. 299, 423-427. - Humble, C., Croft, J., Gerber, A., Ceeper. M., Hame., C. G., and - Tyroler,. H. A. (1990). Passive smoking and 20-yeu cardiovascular ., disease mortality among nonsmoking wivea, Evans county. Georgia. Am. J. Publiclieafth 80, 599-601. Iyengar, S., and Greenhouse, J. B. (1988). Selection models and the Bie-drewetp"roblem. StatrSci 3.; 109-117, , - Jackson, R. T. (1989). Paaslur smoking [Thesis). Aukland Heart Sur- vey. Aukland, New Zealand . Kilpatrick, S. J. (1989). An example of extra-Poisson variation sug- geating aa under-speci8ed model. In Indoor Air Quality (Kasuga, ed). Springer-Verlag, Berlin. Koo, L. C.. Ho, J. H-C., and Rylender, R. (1988). Life-history cone- latea of environmental tobacco smoke: A study on nonsmoking Hong Kong Chinese wives with smoking vetiw nonsmoking hus- bands. Soc. Sci. Med 28, 751-760. Kotelchuk, D. (1974). Asbestos research:. Winning the battle hut los- ing the war. Health Pac Bidt 61, 1-27. LaVecchia. C., D'Avanzo, B., Franzosi, M. C., and Tognoni, G. (1993).

1 I PUBLICATION BIAS IN ETS HEART DISEASE EPIDEMIOLOGY Pass'rve smok)ng and the risk of acute myocardiall infarction [Let- terl.Lancet341,505-506: Layard. M- W. (1995). Ischemic heert disease and spousal smoking in the National Mortwlity Followback Survey. Regul ToxkoL Phcrma- col.21,171-180. Layard. M. W., and Viren, J. R. (1989). Assessing the validity of a Japanese cohort study. In Present and Future of Indoor Air Quality iBieva ec at., eds.). Excerpts Medica. Amsterdam. Lee, P. N. (1992). Enofronmentof To6acco Smuke and Mortatity, Karger AG, 8asel, Switzerland. - Lee. P. N.. Chamberlain, J., and Alderson, M. R. (1986). Relationship of passive smoking to risk of lung cancer and other smoking-asaoci- eud diseases. Br. J. Cancer64, 97-105. Light, R. J. (1987). Accumulating evidence from independent studies: W hat can we win and what oan we lose? Stat. Med. 0, 221-228. Light. R. J., and Pillemer, D. B. (1984). Summing L'p: The Science of Reviewing Research. Harvard Univ. Press, Catqbridge, MA. . ., Martin, M., Hunt, S., and Wllliams, R. (1988). Inereased incidence of heart attacke in nonsmoking women married to smokers. Paper presented at the annual meeting of tbe APHA. In Glantz, S. A., and Parmley, W. W. (1991). Passive smoking and beat diaeax: Epide- miology, physiology and biochemistry. Circulation 83,1-12. .. Sielton, A. W. (1962). Editorial. J. Exp. PsychoL 84, 553-557. Palmer. J. R., Rosenberg, L., and Shapiro, S. (1988). Passive smoking myocardial infarction in women. CVD Neualett. 43, 29. . 191 Peto, J. /1992). Meta-analysia of epidemialogical studies af carcino- genesis. IARC Sci. PubL 110, 571-5 7 7. Rennie, D., and Flanagin, A. (1992). Publication biaa: The triumph of hope over experience. JA.NA 207, 411 i 12. Rosenthal, R. (1979). The "file drawerproblem" and tolerance for null results. PsychoL. Buq. 88, 638-841. Sandler, D. P., and Shore, 0. L. (1989). Quality of data on parents' smoking and drinking provided by adult offspring. Am. J Epide- mio1124,768-i78. - Seeman, I., Poe, G. S., and McLaughlin, J. K. (1989). Design of the 1986 national followback mortality survey: Considerations on col- lecting data on decedents. Pubfic Health Rep. 104, 183-188. Simea, R. J. (198fie). Publication bies: The case for an international registry of clinical triala. J. Clin. Onco6 4, 22-24. Simea, R. J. (1986b). Confionting publication bias: A cohort design for meta-analysia. Stat. Med 8, 11-30. Smith, M. L. (1980). Publication bias and meta-analysis. EouL Educ. 4,22-24. Svendsen. K. H., Kaller, L. H., Martin, M. J., and Ockene, J. K. f1987). Effecta of passive smoking in the multiple risk factor inter vention trial. Am. J. Epidemini 2128, 783-795. Vandenbruucke,JP. (1988). Passive smoking and lungcancer: A pub- lication bias? Br. Med. J. 29@, 391-392. Woolf, B. (1956). On estimating the relation between blood group and disease. Ann, Human C.enet. 19, 251-253.