Tobacco Documents Online

September 28, 1990 Comments by Alan J. Gross, Ph.D. on Chapters 3 and 4 of the EPA Draft Document: Health Effects of Passive Smoking: Assessment of Lung Cancer in Adults and Respiratory Disorders in Children I am a professor in the Department of Biometry, Medical University of South Carolina, Charleston, South Carolina. I received the Ph.D. in statistics from the University of North Carolina in 1962. I have taught at the University of Massachu- setts and the School of Public Health at UCLA. I have also served in the U.S. Public Health Service as a Senior Assistant Scientist attached to the National Cancer Institute. In addition, I am an Associate Editor of the international journal Biometrics and a member of the Epidemiology and Disease Control Stu y Section, National Institutes of Health. My curriculum vitae is attached. In a forthcoming article, presented at the joint 53rd annual meeting of the Institute of Mathematical Statistics and 2nd Bernoulli Society World Congress, Uppsala, Sweden, August 13-18, 1990 and accepted for publication by the Journal of Clinical Epidemiology, Professor Joseph Fleiss of Columbia University and I trace the history and usage of meta-analysis, a statistical tool for combining and integrating the results of independent studies of a given scientific issue, and describe the stringent conditions that must be met for a meta-analysis to be valid. In that article (hereinafter referred to as F-G) and in separate comments submitted by Dr. Fleiss and me on the EPA draft Risk Assessment, we conclude that the meta-analysis conducted by EPA, which forms the heart of the Agency's risk assessment, does not conform to accepted criteria and is invalid. In these comments, I provide a number of more specific observations on the EPA meta-analysis and risk assessment. The principal focus of these comments is on Chapters 3 and 4 of the EPA document. In particular, these comments focus on the EPA assessment of the epidemiologic studies of possible associa- tion between lung cancer and exposure to environmental tobacco smoke (ETS) and the estimates of relative risk from the epidemiologic data. The basic question addressed is whether sufficient evidence exists to establish a relationship between ETS exposure in the general population of the United States and the incidence of lung cancer. To address this question a review and analysis is provided of the existing epidemiologic studies in which individu- als who have presumed higher ETS exposures are compared to those with lower exposures. Typically, the study subjects are married

2 nonsmoking women who are married either to a smoker (presumed higher exposure) or a nonsmoker-(presumed lower exposure). The methodology in the document is a meta-analysis of 21 case-control studies involving married women both exposed and unexposed to ETS. These studies are derived from world-wide literature. As pointed out in the introduction, proper adjustment or control for the biases and confounding that frequently are factors in epidemiologic studies is necessary before these studies can be interpreted properly. As Fleiss and I explain in our paper, the major potential problems with epidemiologic studies include: (i) socio-demographic or clinical differences among study populations, (ii) misclassification of subjects with regard to case-control status and levels of exposure, (iii) confounding factors such as the age and sex of the study sub- jects, and (iv) publication bias in which studies that show no association or a negative association are not published and hence are not included in an analysis of the epidemiologic studies. As shown in these comments, the EPA Risk Assessment is based upon studies for which proper controls were not employed; the meta-analysis of those studies was improper, and the conclu- sion of a causal association between ETS exposure and lung cancer can not•be supported. SPECIFIC COMMENTS ON THE EPA DOCUMENT 1. PP. 1-3, The Document indicts ETS as a Group A (known 1-4: human) carcinogen based on the following: Biological plausibility Consistency of response Upward trend in dose-response Detectable association of environmental exposure levels Broad-based evidence Effects remain after adjustment for potential bias Comment. None of these points has been established. The of biological plausibility is dealt with in other comments. Consistency of response has not been established, as shown by Layard (1990), Letzel et al. (1988) and F-G (1990), among other authors. In particular, F-G establish that there is no statistically significant association between the incidence of lung cancer in nonsmoking females and their exposure to ETS in the U.S., on the basis of a meta-analysis of the U.S. studies, even before adjustment for potential study biases. The purport- ed upward trend in dose-response is based only on the Hirayama (1984) study which has been discredited

3 on multiple grounds by many different authors including Kilpatrick (1987), Lee (1987a, 1987b) and others. If one examines the other large cohort study, Garfinkel (1981), no upward trend is estab- lished. The broad-based evidence purportedly found in the 21 case-control and three prospective studies is suspect on many fundamental levels. It is not at all clear that a statistically significant effect would remain after proper adjustment for bias and confounding. 2. P. 2-4: The document-states: "Of the two major cohort studies, the Japanese study (Hirayama) demonstrates a strong association between passive smoking and lung cancer including an upward trend in dose- response." Comment. This ignores the dose-response information in the other major cohort study, by Garfinkel (1981). The Garfinkel study shows no dose-response relationship. In fact, the relative risk (RR) is 1.27 for women whose husbands smoked between one and 19 cigarettes a day and 1.10 for women whose hus- bands smoked a pack a day or more. (Neither RR is statistically significant.) Moreover, as noted, the Hirayama study has been thoroughly discredited and can not be relied upon. 3. P. 1-5: On this page, the Document reiterates the estimate of 3800 lung cancer deaths per year in nonsmokers attributable to ETS with a 95% confidence interval from 1800 to 6100. The claim is made: "It is unlikely that the number of lung cancer deaths per year attributable to passive smoking by nonsmokers is below 1800..." Comment. In view of the weak and inconsistent results of the reported studies, and the failure to adjust for potential bias and confounding, it is irresponsible to attribute any number of lung cancer deaths in nonsmokers to ETS exposure. 4. PP. 4-1, "After numerically adjusting for background ETS 4-2: (sources other than spousal smoking), the lung cancer risk of ETS from all sources to the U.S. population of nonsmokers (never smokers and former smokers of both sexes) is characterized in terms of lung cancer deaths (LCDS) attributable to ETS (estimated at 3800)." Comments. (i) The U.S. figure is based on a worl wi e estimate. F-G show that if the U.S. studies are considered separately there is no

4 statistically significant association between lung cancer and ETS exposure in the U.S. population, and (ii) the EPA Risk Assessment is unreliable in any event because there is- no accurate method for measuring ETS exposures in nonsmokers. 5. P. 1-3: "The epidemiologic evidence of a lung cancer hazard is statistically assessed by methods of meta-analy- sis to obtain overall results. The data and study results included apply to female married never smokers. Several studies include many subjects, but the percentage of male never-smokers is relatively small and the data are scant by comparison. In some instances, former-smokers are included with never- smokers. All the ETS exposures are considered to be at true environmental levels." Comments. (i) The studies included in the meta-anal- ysis contain many design flaws which are discussed by other commentors, including Peter N. Lee and Maxwell W. Layard. A meta-analysis of flawed studies cannot possibly lead to a scientifically defensible result. (ii) Maxwell Layard as well as this writer have analyzed (by meta-analysis) the results of the male studies. We found a relative risk for the exposed males versus the unexposed males of 1.15 with a 95% confidence interval from 0.61 to 2.16. (iii) Inclusion of former-smokers with never-smokers, how these categories are defined and the attendant misclassification problems all add to the difficulties of drawing conclusions from the existing data. 6. PP. 2-1, 2-2: "Nevertheless, statistical analysis of the combined results could be inconclusive. Statistical significance is evidence that an effect is at a sufficiently high level to be detected with the data available; lack of significance only supports the conclusion that it is below a level that the data have adequately high power to detect with assur- ance." Comment. The quoted statement only discusses one sio~ the issue. Stated in other terms, the statement means that if an effect exists then failure to detect the effect in statistical studies simply means the available data are not of suffi- ciently high power to demonstrate the effect. But what if the effect in question is actually absent? Then not detecting an effect is what should reason- ably be expected from the data. Unfortunately, as

5 indicated by the quoted statement, the Document is premised on the hypothesis that an effect (a relationship between the incidence of lung cancer in nonsmokers and their exposure to ETS) is present and that all one has to do is find adequate data to support this hypothesis. This is an improper reversal of the accepted scientific method. Under the accepted scientific method the investigator begins by assuming the effect (whatever it is) is absent and then attempts to demonstrate by available data that this assumption is incorrect. 7. P. 3-12: "The relative risk comparison of exposed to unexposed individuals, however, is implicitly a comparison of 'exposed to both background and spousal smoke' to 'exposed to background only'." Comment. Although this issue is dealt with in the Document and was dealt with in the NRC (1986) report, the way in which the issue is treated, i.e., assuming that an individual with both background and spousal exposure to ETS has three times the amount of ETS exposure as only background exposure, (NRC report, p. 291) in no way pertains to the individual studies. It is merely a macro-adjustment applied to all studies simultaneously, not a fine tuning of individual studies. This generalized approach is likely to be inaccurate,for individual studies and individual subjects within studies. In many cases-, an individual in the presumed "exposed" category may have less actual exposure to ETS than an individual in the presumed "unexposed" category. For example, the spouse of a smoker may have little ETS exposure from her husband (depending upon when and where the husband smokes) or other sources, whereas the spouse of a nonsmoker may have significant ETS exposure at work or elsewhere outside the home. 8. P. 3-14: In Table 3-5 of the Document, a meta-ana-lysis of the 19 raw studies shows a combined relative risk (RR) for lung cancer among nonsmoking females exposed to ETS (married to a smoking male) of 1.42 compared to nonsmoking females who are unexposed (married to a nonsmoking male). The 95 percent confidence interval for this RR is from 1.24 to 1.63. At the same time, the adjusted meta-analysis in Table 3-6, using those studies showing 95 percent confidence intervals, i.e., omitting the studies of Lam (1985) and Shimizu, et al. (1988) (as the Document does), produces an RR of 1.17 with a 95 percent confidence interval from 0.99 to 1.63.

6 Comment. The latter meta-analysis is consistent wit~ h the null hypothesis that there is no associa- tion between exposure to ETS and lung cancer. Thus, the two meta-analyses produce conflicting results, which undermine the causal conclusions asserted in the Document. 9. P. 3-14: "Data on males is sparse by comparison..." Comment. There are actually six studies on nonsmok- inT g males with lung cancer who were exposed or unexposed to ETS, i.e., the exposed group consists of nonsmoking men married to smoking women and the unexposed group consists of nonsmoking men married to nonsmoking women. Using the notation of the Document, these studies are AK1B, BROW, BUFF, CORR, KABA, and LEE. A meta-analysis of these six studies provides an RR for males of 1.15 with a 95 percent confidence interval from 0.61 to 2.16. The DerSimonian-LaiV test for numerical homogeneity of studies shows X = 3.136 df = 5, p > 0.10, indicat- ing that the male studies are relatively homogenous. One sees that there is no evidence whatsoever that an association exists between ETS exposure and the prevalence of lung cancer in males. 10. On P. 3-21: "Table 3-1 identifies the studies with results adjusted for other variables. Some authors have not included complete details, so the choice of studies for inclusion in this section may be subjective." Comment. Many of the so-called "statistically sig-' ni cant" studies that are included such as AKIB, BUFF, CHAN, CORR, GENG, KABA, KOO, LAMT, AND TRIC, did not adjust for many of the possibly confounding variables.- As discussed in F-G (1990), combining unadjusted studies or studies in which confounding variables are not or cannot be controlled cannot be justified scientifically. 11. P. 3-22: "The Wilcoxon signed-rank test was also applied to the S statistics of Table 3-6 as conducted previous- ly with the raw data, to provide another statistical test of the null hypothesis. The outcome is signif- icant (p = 0.014)." Comment. The combining of the S statistics to o tain a composite statistic is done incorrectly. The reason this approach is incorrect is that all S statistics are given the same weight regardless of sample size. Thus, the S statistic in a small study receives the same weight as the S statistic in a large study. For example, the largest study, Varela

7 (1987) receives the same weight as Inoue and Hirayama (1988) which is a small study. The only correct method of combining study results (if they are combinable, which in this case they are not) is to weight them properly by their size. 12. P. 3-33: "Wells (1989a) reviewed the subject and found it unlikely that publication bias has any substantial effect on the RRs that have been calculated from published reports for passive smoking for either men or women." Comments. Wells' review of the issue of publication bias is inadequate and misleading. The precise extent of publication bias is inherently unknowable. Wells agrees with Vandenbroucke (1988) that data on males is sparse and that there is danger in attempt- ing to extrapolate a male relative risk based on little data. Although Wells attempts to refute Vandenbroucke's arguments in favor of a publication bias, he highlights the weakness of his position when he writes: "If any investigators have data on passive smoking, however, particularly for men, that have not been published or that they have not been able to get published, I would be interested in receiving them for a possible subsequent report." 13. P. 4-2: "Two adjustments are made then to the estimate of RR. The first adjustment accounts for expected bias from former smokers (FS) and current smokers (CS) who may be misclassified as never smokers (NS) and it results in a decrease in the RR estimate. The second adjustment, an upward correction, takes into account the risk from background exposure to ETS (experienced by a NS whether married to a smoker or not)." Comments. (i) There is no clear definitional dis- tinction between NS and FS. Should someone who smoked one year and quit be lumped with FS or NS? There are problems either way. (ii) Background exposure is difficult to measure. Thus, the assess- ment of exposure to spousal smoking plus background as compared to background exposure alone is merely speculation. It then becomes an exercise in formula manipulation to place numerical values on ETS exposure levels. While it is probably not unreason- OD able to assume that most of the population is I exposed to some background level of ETS, this level N will vary between and among communities in the U.S. Cn The EPA assumption that ETS exposure for spouses of f4 smokers is three times background exposure is a 0 misleading generalization. In fact, on p. C-17, the CO

8 Document concludes "Exposure to ETS may vary widely due to differences in cigarette type, rate of smoking, ventilation conditions, room volume, etc. No attempt is made here to develop calculations under the immense range of conditions likely to be found in society. Instead calculations are presented for a simplified case that is typical of exposure conditions." 14. P. 4-3: "The population attributable risk then is estimated to obtain an annual number of U.S. LCDs in never- smoking women attributable to total ETS exposure. The predicted number of LCDs due to ETS is further extended to include male NS and then to include former smokers of both sexes." Comment. Such an extension is based on assumptions that cannot be justified. As noted, available data show no association between male NS exposure to ETS and lung cancer incidence. 15. P. 4-13: The document includes an assessment of lung cancer risk from ETS exposure based on comparisons of Seventh-Day Adventists (SDA) and non-SDA popula- tions. "The comparison of two groups of NS is based on the premise that the non-SDA cohort is more likely to be exposed to ETS than the SDA groups due to differences in life-style." Comment. There are other profound differences in lifestyle between SDA and non-SDA populations besides exposure to ETS and unless those differences are properly adjusted for it is impossible to isolate any particular effect of ETS. (i) The difference in lung cancer deaths between the groups may well be due to any number of causes other than ETS exposure. SDA's and non-SDA's tend to differ in major lifestyle variables, such as alcohol consump- tion, occupation and dietary habits, and racial and ethnic differences; (ii) selection bias is possible since SDA participants were selected from active church members whereas the other non-smoking partic- ipants were part of an American Cancer Society cohort which was selected from among friends, relatives and acquaintances of American Cancer Society volunteers; (iii) among SDA members, the causes of about 10% of the out-of-state deaths (the study took place in California) were not ascer- tained. There is also an underascertainment of causes of death among older subjects.

9 16. P. 4-23: "Although it is likely that the true differential in ETS exposure for subjects classified as exposed or unexposed is relatively higher in some sampled human environments than others (probably higher in Japan, for example, as discussed by Hirayama and others), comparison of outcomes with study characteristics did not reveal any apparent patterns or study characteristics associated with the findings." Comments. This justification falls far short of what is required before a proper meta-analysis can be conducted. This is discussed in detail in F-G. 17. P. 4-33: "The true values of RR being estimated depend on both the study design and protocol. Culture, environment, and lifestyle would influence inter- study differences. In particular, one might expect these factors to contribute to intercountry vari- ability in the epidemiologic data. To extract this source of variability statistically would require multiple studies from several countries among those analyzed (China [2j, Hong Kong (4), Japan (2), Sweden [2), U.S. (8], the studies are not suffi- ciently similar within countries to test variabili- ty. In particular, there is considerable dissimi- larity between the U.S. studies, and this probably contributes to their wide ranging results." Comment. This statement provides a compelling reason not to engage in a meta-analysis of the ETS studies. It is precisely the reason that meta-analysis ordinarily should not be used to combine heteroge- neous studies. Reiterating F-G, "Meta-analysis, a set of statistical tools for combining and integrat- ing the results of independent studies of a given scientific issue, can be useful when the stringent conditions under which such integration is valid are met." In the above quoted statement, EPA concedes that these considerations have not been met.

10 REFERENCES Akiba, S.: Kato, H.; Blot, W.J. (1986) Passive smoking and lung cancer among Japanese women. Cancer Research 46: 4804-4807. Brownson, R.C., J.S.; Keefe, T.J.; Ferguson, S.W.; Pritzl, J.A. (1987) Risk factors for adenocarcinoma of the lung. Am J Epidemlol 125:25-34. Buffler, P.A.; Pickle, L.W.; Mason, T.J.; and Contant, C. (1984) The causes of lung cancer in Texas. Mizell, M. and Correa, P., eds. Lung Cancer: Causes and Prevention. New York: Verlag Chemie International, pp. 83-99. Chan, W.C.; Fung, S.C. 1982) Lung cancer in non-smokers in Hong Kong. Cancer Campaign, Vol. 6, Cancer Epidemiology, (Grundmann, E., ed.) Stuttgart: Gustav Fisher Verlag, pp. 199-202. Correa, P.; Fontham, E.; PicKle, L.; Lin, Y.; Haenszel, W. (1983) Passive smoking and lung cancer. Lancet 2:595-597. Cummings, K.M.; Markello, S.J.; Mahoney, M.C.; Marshall, J.R. (1989a) Measurement of lifetime exposure to passive smoke. Am J Epidemiol 30:122. DerSimonian, R.; Laird, N. (1986) Meta-analysis in clinical trials. Controlled Clin Trials 7:177-188. Environmental Protection Agency (1990) Health Effects of Passive Smoking: Assessment of Lung Cancer in Adults and Respiratory Disorders in Children. Draft Document Fleiss, J.L.; Gross, A.J. (1990) Meta-analysis in epidemiology, with special reference to studies of the association between exposure to environmental tobacco smoke and lung cancer. A critique. (Accepted for publication by the Journal of Clinical Epidemiology.) m ~ C~rt Cri iV M+ N