Stanton Glantz' Post-OSHA Hearings Comments (1996)


POST HEARING COMMENT

Stanton A. Glantz, PhD

Credibility, Causality, and Other Word Games

As will be discussed in this Post Hearing Comment, the tobacco industry presented a large number of witnesses who raised many, often inconsistent, criticisms of OSHA's Notice of Proposed Rulemaking and the science upon which it is based. There are, however, some important patterns that emerge from reviewing the record that bear comment.

First, the tobacco industry witnesses limit themselves to detailed (often inconsequential) criticisms of the individual research projects reported in the scientific literature, while steadfastly refusing to consider the evidence as a whole.

Second, they fail to apply the same criticisms or level of scrutiny to work that supports their position as work that supports the conclusion that ETS causes disease.

Third, they establish a standard of proof that is impossible to meet, such as suggesting that the only way to be sure of a conclusion is to conduct an experiment such as a randomized controlled trial. It is simply ethically and practically impossible to randomly assign people to be exposed to controlled levels of ETS for a given length of time, wait for them to die (or kill them), then conduct autopsy to verify the precise cause of death. Animal experiments, in which one could control these variables, are rejected because animals are not people. To meet the objections that the tobacco industry witnesses raised to the evidence OSHA and others have developed that ETS causes disease one would have to do a human study. Even so, many of the same objections raised in the record could be raised: Were the subjects in the sample typical of the American population? The working population? Were the exposure chambers similar to workplaces? Was the controlled diet similar to the diet eaten by all American workers? The list goes on and on.

Fourth, and most important, they deny the obvious. Of the 17 witnesses appearing at the request of the tobacco industry who were asked whether active smoking causes lung cancer or other diseases, not one simply answered "yes " (Holan, p. 4359; Idle, p. 5201; Hubert, p. 5456- 5457; Ashford, p. 5682-5683, 5685; LeVois, p. 5807-5809; R. Witorsch, p. 6089-6091; Springall, p. 6188; Switzer, p. 6335-6340; Starr, p. 6945; Layard, p. 6695; Roth, p. 9192-9195; Gori, p. 9562-9565; Bridges, p. 10088-10093; Holcomb, p. 10776, 10779-10780; P. Witorsch, p. 10643; Newell, p. 10985; Coggins, p. 11633; the actual questions and answers are in Attachment 1). Some did not know, some said it was beyond their professional expertise, some said active smoking was a "risk factor" (the current term the tobacco companies use). Despite persistent questioning, only one witness (Hubert, p. 5456-5457) finally admitted that active smoking causes disease. Indeed, LeVois stated that "you cannot prove causality" even in a controlled experiment (p. 5901).

This situation contrasts with other carcinogens unrelated to tobacco. When asked if anything had been found that causes lung cancer in people, RJ Reynolds' Coggins promptly and unambiguously answered "asbestos," (p. 11641) despite the fact that the precise mechanism by which asbestos causes cancer is not precisely understood. Indeed, there is probably a better understanding of how tobacco causes cancer than how asbestos causes cancer. In any event, despite the fact that the nature and type of evidence available in both cases is similar, Coggins had no trouble concluding that asbestos causes cancer whereas he found the evidence on smoking inadequate.

One can only speculate why the tobacco industry and its lawyers are willing to say "risk factor" but not "cause," but it is clear that the tobacco industry's witnesses avoided the word cause the way a vampire avoids light. This refusal to accept probably the best established fact in biomedical science draws everything else these people have say into question.

Publication Bias

One central argument raised by the tobacco industry and its witnesses is that OSHA (and others) have concluded that the reason that there so many scientific bodies have concluded that ETS causes lung cancer, heart disease, and other problems is the presence of publication bias, wherein studies that fail to find that ETS causes diseases are not published. Several reasons for this failure to publish were given: (1) bias of editors against "negative results," (2) failure of scientists to submit papers that do not show that ETS causes disease, and (3) failure of people to conduct studies that ended up yielding negative results, particularly the National Death Followback Survey and the American Cancer Society CPS-I and CPS-II data sets. Several of the tobacco industry witnesses used "funnel plots" to bolster their arguments that such a publication bias exists in the case of ETS and heart disease.

The argument that publication bias exists -- particularly when it is defined as broadly as the tobacco industry witnesses defined it -- is a particularly convenient argument to raise because it is an argument about what is not there. The nonexistent studies cannot be critiqued for their methods or related to the other studies.

There are several serious problems with the industry's arguments:

There have been several attempts to address the question of publication bias against negative studies in the ETS literature, all of which have concluded that there is no such bias.

Bero et al [1] examined the available literature and searched for unpublished data sets on ETS and lung cancer and failed to find any evidence of publication bias. Indeed, if publication bias is defined as failure to publish small studies that do not show a statistically significant effect (in this case, of ETS on lung cancer), they presented evidence that there is a bias in favor of publication of negative or equivocal studies on ETS.

Beaglehole [2] did a formal search for unpublished data sets and studies on ETS and heart disease and could not locate any.

Wells [3] noted that the few abstracts on ETS and heart disease that did not ultimately appear as full papers all reported an increase in risk of heart disease associated with ETS exposure.

Finally, Vandenbroucke [4], originator of the "funnel plot" that several tobacco industry witnesses used to argue that there is a publication bias against negative papers on ETS and heart disease, concluded that no such publication bias existed for ETS and lung cancer in women. It is interesting that none of the tobacco industry witnesses, who discussed funnel plots at such length, pointed out Vandenbroucke's conclusion.

The Vandenbroucke "funnel plot" has never been shown to demonstrate that it actually detects publication bias. In addition, this technique was not consistently applied or interpreted by the tobacco industry's witnesses.

It is important to emphasize that the "funnel plot" is simply an interesting suggestion as a way to think about publication bias. The original Vandenbroucke paper [4] is only 1_ pages long and contains no empirical evidence that it proves publication bias exists. When asked for such evidence, none of the tobacco industry witnesses could point to any empirical evidence that a funnel plot actually detected publication bias. In addition, Vandenbroucke did not propose any formal analysis of funnel plots to quantify publication bias.

LeVois applied the funnel plot more-or-less as Vandenbroucke originally proposed, with the exception that he added a formal test to see if the data followed the funnel pattern. He did so by applying a Pearson product moment correlation (p. 5884-5885) to the log relative risk versus standard error of the log of the relative risk. When questioned about this procedure, LeVois saw no problems with this approach. There is, however, a fundamental difficulty with what LeVois did. The Pearson product movement correlation coefficient is a parametric statistical test that requires making several assumptions, most notably that the variance in the observations is constant, i.e., the data are not displayed in a "funnel pattern." Virtually any introductory statistics textbook that discusses Pearson correlation in any depth makes this point. Thus, the statistic that LeVois computed is simply not appropriate for analyzing the data he presented.

One could compute the correlation among the points that LeVois presented using a Spearman correlation -- which does not require making the constant variance assumption. Doing so fails to yield a statistically significant P value. Thus, if one accepts the general analytical approach advanced by LeVois (which I do not) and simply uses the correct test statistic, LeVois' data yields the conclusion that there is no publication bias.

In comparison to use of the Pearson correlation in an ad hoc attempt to be quantitative about the analysis of funnel plots as LeVois did, Springall, another tobacco industry witness, specifically eschewed any formal analysis of the plots and chose to rely on subjective judgement (p. 6448-6449, 6453). In fact, the decision that the data are displayed in a "funnel" is an arbitrary and subjective decision. (Vandenbroucke also draws his conclusions by simply looking at the plot.) Indeed, both LeVois (p. 5884-5885) and Springall (p. 6448-6449) simply drew the funnel lines on the plots rather than computing them from the data using any objective procedure.

Later, when asked about LeVois' procedure of computing the Pearson correlation, Springall changed his view and said it was an acceptable procedure (p. 6457-6460). Springall stated that it was reasonable to do such a computation after looking at the data and seeing that it appeared to fall on a straight line. This kind of post hoc analysis was soundly criticized by another tobacco industry witness, Switzer, who said (p. 6251),

Unless researchers state clearly in advance of the data collection, and I emphasize in advance, exactly how they will report the results, there is a real possibility that groupings and other adjustments may be chosen which more clearly exhibit positive findings. I think it's perhaps a natural instinct on the part of people who do studies to try to find positive findings in their research and therefore they may make choices after seeing the data. I would be surprised if they didn't.

Switzer, however, did not see fit to raise the same criticism about LeVois' post hoc analysis of the funnel plot.

The tobacco industry's witnesses were inconsistent in their use of the funnel plot; Springall did not apply the funnel plot methodology the same way LeVois did. Rather than plotting the standard error of the logarithm of the relative risk on the horizontal axis (as LeVois and Vandenbroucke did), Springall plotted the square root of the sample size on the horizontal axis. When questioned on why he did not construct the plots as Vandenbroucke originally proposed, he said that he did it because the standard errors were "crazy" (p. 6444). Leaving aside the question of the precise scientific definition of a "crazy" standard error, it is worth noting that simply using the square root of the sample size means that the value along the horizontal axis simply reflects how big the study is, without adjusting for any potential confounding variables. Given the extensive discussion of the importance of confounding variables that the tobacco industry witnesses presented -- including Springall -- it is surprising that Springall chose to ignore any effects of confounding in his analysis of the funnel plot. This point is another example of the inconsistent methods and standards applied by the tobacco industry witnesses.

Most of the papers that have cited Vandenbroucke's paper conclude that ETS causes disease. A search of Science Citation Index located 35 papers, 14 of which dealt with ETS (see Attachment 2). Of these 14 papers, 12 (86%) concluded that ETS harmed health. Both of the papers that stated that there were not enough data to reach such a conclusion were written by scientists with financial connections to the tobacco industry; one of them was by LeVois and Layard [5] and essentially repeated the testimony presented to OSHA and discussed in this post hearing comment and the other was prepared by Fleiss and Gross [6] at the suggestion of Myron Weinberg of WASHTEC (a tobacco industry witness at this hearing) with funding from the Tobacco Institute.

Despite the fact that several of the tobacco industry witnesses claimed that there was a bias against publishing -- or even conducting -- negative studies, none of them could point to a specific instance in which they had this problem (Idle, p. 5220; Newell, p. 11052; Coggins, p. 11634). The only specific case that was discussed was an attempt by Ashford to publish a study in the New England Journal of Medicine indicating that children of smoking mothers had better survival rates than children of nonsmokers. The New England Journal rejected this study, but Ashford was able to publish it in the Italian Journal of Obstetrics and Gynecology (p. 5770- 5771). He was unable to locate the reviews of the paper or provide any evidence that the New England Journal had not simply rejected the study on the grounds of poor quality. Moreover, the fact that the paper was ultimately published (albeit in a less prestigious journal) is evidence against the existence of publication bias.

Several tobacco industry witnesses made an issue of the fact that the American Cancer Society had never published an analysis of the CPS-I and CPS-II data sets that examined the question of passive smoking and heart disease and cited this as evidence of publication bias (e.g, Ashford, p. 5654; LeVois, 5768; Gori, P. 9542; Sears, p. 11993; Lee's written submissions, Sirridge's cross examination of Glantz, p. 578- 580). The fact that scientists had not gotten around to doing a specific study is not evidence of publication bias. It is, after all, the American Cancer Society, and heart disease is not a priority for this organization.

In response to these criticisms, I contacted Dr. Michael Thun, Director of Analytical Epidemiology and the American Cancer Society and asked him if there were any results available on ETS and heart disease from these studies (Attachment 3). He informed me that, while they were interested in this question, it had been a low priority for the American Cancer Society. He was kind enough to conduct a preliminary analysis of the data which, in contrast to the analysis presented at the Hearing by LeVois and Layard, showed an elevation in risk of heart disease for nonsmoking men married to women who smoked with a relative risk of 1.21. This value is consistent with the other studies in the literature. (Dr. Thun is now conducting a more complete analysis of this dataset, which will hopefully become available in the near future.)

There are very serious problems with the analysis of the CPS-I and CPS-II data sets put in to the record by LeVois and Layard; their failure to properly account for important confounding variables, particularly age, as well as the likelihood of serious misclassification of nonsmokers as smokers, biases their results towards the null, i.e., missing a real effect of ETS on heart disease.

As Michael Thun noted when I asked him to review LeVois and Layard's written submission:

Although LeVois and Layard reportedly controlled for age in the Poission regression analysis (Table 4), they do not present age-specific data. Active cigarette smoking is known to be a stronger risk factor for CHD death at younger than at older ages. For example, among white male active cigarette smokers in CPS-II, the RR for fatal CHD (compared to never-smokers) decreases from about 6.3 at age 40-44 to 1.4 at ages Æ75. Failure to examine modification of the association by age could easily obscure a true association between passive smoking and CHD mortality.

Perhaps it is this failure to properly account for the confounding variable of age that led LeVois and Layard to fail to detect that effect of ETS that appeared in Thun's preliminary analysis. LeVois was extensively questioned on the relationship between age, smoking, and heart disease risk by Mr. Myers (p. 6008-6011) and failed to recognize the importance of this confounding variable or how to properly correct for it in his analysis.

There is another, more fundamental, problem with the LeVois and Layard analysis of the CPS data sets: the numbers do not add up correctly. As Thun's letter notes, Table 2 of LeVois and Layard's written submission stated that there were 226,067 never-smoking women in CPS-II. This exceeds by over 13,000 the number of women included in a parallel American Cancer Society Study of lung cancer and ETS. Despite extensive questioning on this discrepancy, LeVois could not explain this error of 13,000 women (p. 5861-5867, 6004-6005). During cross examination by Mr. Myers, LeVois admitted that their analysis counted any individual who was married to an ever-smoker as exposed to ETS, even if the ever-smoker quit 10-15 years earlier, perhaps before marrying the person who was the case in the CPS data set (p. 6005-6008). Given the fact that the risk of heart disease in active smokers declines quickly upon smoking cessation, falling by 50% in one year, and essentially back to that of a nonsmoker over a few years [7]. Since the mechanisms by which active and passive smoking cause heart disease are similar, it is likely that ending exposure to ETS will be followed by a similar fall in the risk of heart disease associated with past ETS exposure. As a result, LeVois' practice of including ever smokers as markers of ETS exposure seriously biases his results towards the null, i.e., his analysis will systematically underestimate the risk of current ETS exposure as a cause of heart disease.

It is also worth noting that neither the CPS-I nor CPS-II studies were primarily designed to study the effects of ETS on heart disease. As discussed by Wells in his written submission and testimony (p. 965- 967), the questions used to assess ETS exposure in these studies are limited and raise the possibility of serious biases towards the null in the analysis of the effects of ETS on lung cancer and heart disease. Even given these biases, it is notable that proper analyses of these data sets shows an increase in risk of both lung cancer and heart disease associated with ETS exposure.

Given the extensive critique of the possibility of misclassification of smokers as nonsmokers that several of the tobacco industry witnesses, it is remarkable that none of them raised the issue of misclassification in connection with this study, where it is likely to be a much bigger problem because of the large number of former smokers.

These same criticisms apply to the publication by LeVois and Layard in Regulatory Toxicology and Pharmacology [5] that essentially repeats the written materials submitted to OSHA. Interestingly, in his oral testimony on November 15, 1994, LeVois said "These data are presented to OSHA in an appendix and are presently being prepared for submission to a journal for publication." In contrast, the published article in the journal states that it was received for publication on June 11, 1994. It is also noteworthy that Gio Gori, another tobacco industry witness, is an associate editor of the journal.

The fact that the analysis of the National Mortality Followback Survey yielded negative results is neither evidence that ETS does not cause heart disease nor that there is a publication bias against publication of negative studies because this analysis has so many problems as to be worthless.

The fundamental problem with this analysis is the definition of a smoker. Someone is identified as a smoker if he or she ever smoked 100 cigarettes in their entire lives. While this is a reasonable definition of never smoker for the cases in the studies, this same definition was used for the spouses that were supposedly exposing (or not exposing) the cases to ETS. Thus, a spouse who was 60 years old but who had smoked only as a teenager would be counted as a smoker and the case would be counted as "exposed." Given the tens of millions of former smokers, this situation created huge misclassification errors in who people are counted as exposed to ETS when, in reality, they are married to a nonsmoker. As of 1987, 40.3% of ever smokers had been nonsmokers for at least one year [7]. This misclassification error strongly biases the results towards the null, i.e., a relative risk of 1. It is remarkable that, at the same time that several tobacco industry witnesses provided extensive discussions of the effects of relatively small misclassification errors (of smokers counted as nonsmokers) in the published studies, not one of them raised this issue in the Layard study of the National Mortality Followback Survey. Despite extensive questioning on this point, Layard showed no evidence of having even considered this problem (p. 6655-6661).

This study is also an example of the inconsistent application of scientific standards by the tobacco industry witnesses to studies that support and oppose the industry position that ETS is not dangerous. While the studies in the peer reviewed literature that support a link between ETS and disease were soundly criticized (particularly the Fontham study) for using surrogate responses, the National Death Followback Survey relied entirely on surrogate responses. None of the tobacco industry witnesses included this problem in their discussion of problems with epidemiological studies on ETS and disease.

These same criticisms apply to the publication by Layard in Regulatory Toxocology and Pharmacology [8] that essentially repeats the written materials submitted to OSHA. Interestingly, in his oral testimony on November 17, 1994, in response to whether the work was published in a peer reviewed journal, Layard said "it's about to be submitted." In contrast, the published article in the journal states that it was received for publication on June 25, 1994. It is also noteworthy that Gio Gori, another tobacco industry witness, is an associate editor of the journal.

Confounding Variables

In addition to publication bias, the tobacco industry witnesses spoke at length of the problem of confounding of ETS exposure with other risk factors for heart disease, lung cancer, and other disease caused by ETS. Samet, one of OSHA's witnesses, provided a clear and detailed discussion of confounding variables when he was cross examined by Messrs. Grossman, Sirridge, and Rupp (p. 192-296). While the tobacco industry's lawyers concentrated on getting Samet to agree that confounding could make it appear that ETS caused disease when it did not, Samet correctly noted that confounding variables can also mask an effect when one exists. Which situation actually exists depends on the specific case at hand.

Here is how confounding variables can make it appear that something causes a disease when, in fact it does not: If A causes B and A causes C, B and C will tend to change together (in response to A), even if B does not cause C because of the effect of the confounding variable A. This situation, in which B is "unjustly" concluded to cause C has led the tobacco industry witnesses to a long list of possible confounding variables that can be used to explain the observed associations between ETS exposure and heart disease and lung cancer. Some of these potential confounding variables include race, age, diet, socioeconomic status, and pet birds. The tobacco industry's witnesses spoke at great length about these possibilities (e.g., LeVois, p. 5771- 5775, 5778, 5782, 5786, 5797, 5805-9, 5814-5815, 5818, 5835, 5841-46, 5850; Witorshch, 6117-6119, 6141-6144; Springall, p. 6173-6175, 6177, 6180, 6183, 6229-6231; Switzer, p. 6247, 6229-6301, 6306, 6310-6314, 6317-6318; Starr, p. 6495-6496, 6502, 6530, 6540; Layard, p. 6592, 6594, 6595-6597, 6601-6604, 6613-6614; Roth, p. 9091-9095, 9098-9910, 9113- 9115, 9128, 9131, 9133, 9154-9155, 9193, 9202, 9210-9215, 9220, 9224, 9227-9233, 9245-9247, 9251-9252, 9256-9257; Coggins, p. 11450-11451, 11466, 11467-114672, 114679, 114682-114683, 114696, 11563). In addition, confounding was the subject of extensive cross examination of OSHA witnesses by the tobacco industry's lawyers (Glantz, p. 531, 541, 561, 579, 731-732, 739; Samet, p. 1192-1196; Benowitz, p. 1233, 1237, 1239, 1279-1280, 1346-1347, 1361-1362; Steenland, p. 1936-1939, 1944, 1950, 1960-1963, 1988-1984, 1990, 1993-1994, 1996, 2078-2081, 2086- 2092.) Despite this extensive presentation, and specific questions from OSHA during cross examination, none of the tobacco industry witnesses provided any empirical evidence that the potential problem of confounding really existed.

Any good epidemiologist should consider possible confounding variables when interpreting the results of an epidemiological study. The fact that someone can list a variable as a potential confounder, however, does not constitute evidence that it is a confounder. (For example, both the EPA [9] and Fontham et al [10] carefully evaluated diet as a potential confounder for ETS and lung cancer and found that the ETS effect on lung cancer persisted even after controlling for diet. Hence, any dietary differences between people exposed to ETS and not exposed do not explain the increased risk of lung cancer death associated with ETS.) Indeed, the tobacco industry's witnesses' presentations simply consisted of lists of potential confounding without any empirical evidence or quantitative analysis to support the assertion that these confounding variables actually explained the observed associations between ETS and disease or that such confounding effects, even if present, were large enough to explain the observed associations between ETS exposure and disease. The two exceptions to this failure to present quantitative evidence that confounding, if present, could explain the association between ETS exposure and disease were Springall (p. 6180) and Roth (p. 9133, 9256-9257). In both cases, the witnesses assumed an effect of confounding and then analyzed the impact of this assumed effect. When pressed for the empirical evidence that the effects were as assumed, none was produced. Nothing in the massive amount of material submitted in writing and as testimony at the Hearing by the tobacco industry and its witnesses, constitutes affirmative empirical evidence that real confounders existed that could explain the observed link between ETS exposure and disease. Indeed, the failure of the tobacco industry to produce evidence that real confounders exist increases the confidence that OSHA can have in concluding that ETS causes disease.

As explained by Samet in his cross examination by tobacco industry lawyers, confounding can also mask a real risk. The only industry witness to recognize that it is possible to "overcorrect" for confounders was Starr (p. 6495-6496), who only recognized this fact during cross examination. Any good textbook on multivariate analysis cautions against this procedure; including redundant variables creates a situation known as "multicollinearity" in which real effects of the independent variables are masked because the uncertainty in the parameter estimates (i.e., estimates of the effects of the variables, such as ETS exposure, on the outcome variable, i.e., death). In contrast, when questioned on whether it is possible to "overcorrect" by including too many independent variables in an analysis to evaluate the effects of ETS, LeVois incorrectly stated that it was a good thing to have overlapping independent variables in the analysis (p. 5925-5927).

Likewise, Springall incorrectly stated that multicollinearity increases the risk estimates (p. 6430-6432). Of all the tobacco industry witnesses, in response to cross examination, only Switzer (p. 6324-6328) correctly described the fact that multicollinearity could mask a real effect. Switzer did not, however, raise this issue in criticizing the long lists of overlapping potential confounding variables suggested by other tobacco industry witnesses.

Lee, another tobacco industry consultant, provided an extensive discussion of confounders in his written submission, but did not appear at the hearing to answer questions about his work. In particular, he suggested 33 confounders for heart disease that he stated should be considered in an analysis of the effects of ETS. As noted in the attached tables (Attachment 4), many of these confounders are multiple measurements of the same thing (e.g., 11 dietary variables plus overweight, underweight, and fatty foods). Including all these confounders would create serious multicollinearity problems which could make it impossible to detect a real ETS effect even if one is present. Ironically, LeVois, who spoke at length about the need to account for confounding variables in the analysis of ETS, failed to account for most of Lee's variables in his analysis of the CPS data sets (p. 5919-5925). LeVois even call Lee's list "odd" (p. 5920) when asked one-by-one about Lee's potential confounders without knowing the source. (Springall also questioned the wisdom of including several variables that measured the same thing; p. 6436-6437.) Later, Mr. Rupp submitted LeVois to extensive friendly cross examination to justify LeVois' failure to control for all of Lee's confounding variables (p. 6031-6033, 6039-6041, 6049-6051, 6267, 6270-6274).

Similar to LeVois, Layard's analysis of the National Death Followback Survey did not take into account 100 potential confounding variables Lee suggested for lung cancer (p. 6651-6655). Later, Mr. Rupp offered Layard friendly cross examination (p. 6701-6702) to justify the fact that Layard did not correct for all these confounders, primarily because Layard only had a small number of cases and, as a practical matter, doing such corrections was impossible. It is interesting that