Passive Smoking: Introduction

Achtung! If you're going to read this section (and, of course, you should), you will find it much easier as a PDF. This is the complete Appendix of Velvet Glove, Iron Fist. View it here.

The Evidence

Glossary of terms


"There is overwhelming evidence, built up over decades, that passive smoking causes lung cancer"

- Vivienne Nathanson, British Medical Association

Takeshi Hirayama conducted the first epidemiological research into passive smoking by monitoring the health of nonsmoking women married to smoking husbands and this model remains the gold standard for research of this kind. In the 25 years since Hirayama's paper was published (1981), a further 62 similar reports have been published. Taken together they form a substantial body of evidence which, according to one Surgeon General, is 'overwhelming' in supporting the hypothesis that nonsmokers exposed to secondhand smoke are more likely to suffer from lung cancer than those who generally avoid exposure. After reading all of these studies,, I have not been able to endorse this interpretation.

There are only three possible outcomes in studies of this kind. The first is that the hypothesis is correct (ie. that passive exposure to tobacco smoke increases lung cancer risk). The second possibility is that there is a negative association (ie. that passive smoking reduces the risk of lung cancer). The third possibility is that there is no association either way; this is known as the 'null hypothesis'.

A relative risk (RR) of 1.0 represents no association either way. An RR below 1.0 represents a negative association and an RR above 1.0 represents a positive association (increased risk). For example, 0.9 = 10% less risk, 1.35 = 35% greater risk and 2.0 = 100% greater risk.

In the studies below, researchers typically compare a group of married, nonsmoking women who have lung cancer (the cases) with a group of married, nonsmoking women who do not (the controls). Questions are asked of both groups regarding exposure to tobacco smoke and, to put it in simple terms, if 60% of the lung cancer cases are married to smokers and only 40% of the controls are married to smokers, it might be inferred that marriage to a smoker increases lung cancer risk by half (RR = 1.50).

Or it might not. In this example, can we be sure that it was the husband's smoking habits that made the difference? Can we be sure that the cases were not smokers themselves? What if the cases all came from the city or all worked in the asbestos industry? What if they are older than the controls and therefore have a higher risk of cancer anyway. Are their diets comparable? Did they used to be smokers? All these are confounding factors and need to be identified and avoided. If they are unavoidable, the figures must be adjusted to take them into account.

The numbers involved are crucial if we are to draw any conclusions from a study of this kind. In the example above, the women married to smokers appear to have a relative risk of lung cancer of 1.5; a 50% rise. That is based on 40% of the healthy controls being married to a smoker but what if there were only 5 women in each group? That would mean two of the controls were married to smokers and three were not. Among the cases, this ratio is reversed. Technically, 40% of the controls were married to smokers but does that mean that 40% of the general population are married to smokers? The difference between the two groups comes down to one woman in each group saying that she is or is not married to a smoker. The study is vulnerable to because of its small size and may not be - indeed, probably is not - representative of the population at large. If, on the other hand, 1,000 women are in each group and the percentages remain the same we can say with rather more confidence that passive smoking increases lung cancer risk and that this lies somewhere around 1.5.

Clearly, we must exercise extreme caution before drawing conclusions from small sample groups but as the number of participants increases, the margin for error is reduced and our estimates should become more accurate. As discussed in chapter 7, epidemiologists distinguish between chance results and genuine associations by using a standard of statistical significance. In our example of 5 lung cancer patients, the RR is 1.5 but this does not tell the whole story. The full RR is 1.5 (0.4-6.0) with the figures in brackets being the lower and upper limit. Because of the small number of cases, the confidence interval (or the margin of error) is very wide and we can only surmise that the risk to the nonsmokers falls between a 60% reduction and a six-fold increase - not very useful.

In the second example, because there are 1,000 cases, the margin of error is much narrower and the RR is 1.5 (1.1-2.0). This tells us that risk is increased by at least 10% and may be as high as 100%. This association is statistically significant because the lower limit of the confidence interval is above one. If it was 1.0 or lower it would not be, since the RR includes the null hypothesis and the negative hypothesis. And if an RR is not statistically significant, it tells us nothing. It does not matter whether the headline figure is higher or lower than 1.0, it supports neither the positive or negative hypothesis. The null hypothesis itself cannot be proven even in the unlikely event of the RR landing exactly on 1.0. However, if enough studies show nonsignificant findings, one might reasonably infer that there is no association to investigate.

Of the epidemiological papers that studied the effect of secondhand smoke on nonsmoking wives, 9 found a statistically significant positive association, 2 found a statistically significant negative association and the remaining 52 found no statistically significant association either way. Some within the tobacco control movement have claimed that the risk from passive smoking is too small to be demonstrated conclusively in small and medium sized studies. Only very large studies, they say, have the statistical power to meet the criteria for significance but these studies are difficult to carry out thanks, in part, to the relative scarcity of lung cancer patients who have never smoked. There is some truth in this, although it is worth pointing to the 11 findings here that have achieved statistical significance.

Since the mid-1980s, it has become clear that the early reports from Hirayama and Trichopoulos that showed a doubling of lung cancer risk were erroneous and that if a risk exists at all, it falls at a level below 1.50. In the past twenty years, very few statistically significant associations have come to light and so those who have put their faith in the passive smoking theory have used the nonsignificant findings found in the bulk of studies to make their case. They have claimed that although the majority of epidemiological papers do not show significant associations, the weight of evidence points towards a positive association and that, taken together, they show a risk of around 1.25.

While it is unusual to infer anything from relative risks that do not meet the minimum standard of statistical significance, it is not completely unreasonable to draw conclusions if they all point in the same direction and show a similar relative risk. However, that is certainly not the case here.

Of the 52 statistically insignificant results, 20 have a relative risk of 1.0 or below and 32 have a relative risk above 1.0. The best that can therefore be said of this data is that there are more studies pointing towards a positive association than a negative one. This is feeble stuff. No one is claiming that secondhand smoke protects people from lung cancer but if the 20 studies that point in that direction are not to be trusted, why trust the 32 that point the opposite way?

If we accept that secondhand smoke causes lung cancer in nonsmoking women because two-thirds of the nonsignificant results lean that way then we must also accept that women who are exposed to secondhand smoke in childhood are less likely to suffer from lung cancer (two thirds of the studies regarding passive smoking in childhood have shown a negative correlation).

If the majority of studies showed relative risks that were closely grouped between 1.20 and 1.30 that one might be more inclined to accept the plausibility of the passive smoking hypothesis. Britain's SCOTH committee and anti-smoking groups around the world have now settled on a relative risk for secondhand smoke and lung cancer of 1.24 but not one of the 63 studies below shows a risk of that magnitude and even if one allows a generous margin of error and settles for any risk between 1.15 and 1.35, there are only ten studies that fit the bill.

For every study that shows a statistically significant positive association, there are six that do not. This is hardly overwhelming evidence in support of the passive smoking theory and yet these nine significant associations do exist, compared to 'only' two in the opposite direction. Are they suggestive? The reader should not infer that it is difficult or unusual for a random result to achieve statistical significance and should remember that for every significant finding shown below, there are six that support the null hypothesis. Most, if not all, of the transient health scares covered in Appendix B reached this minimum scientific criteria before making it into the newspapers, but this was no guarantee of veracity. Any epidemiologist who asks questions about enough aspects of their subjects' lifestyle will chance upon plenty of apparent associations and although passive smoking may seem a limited field there is plenty of scope for data-dredging.

A study of one case group can, therefore, produce over a hundred individual risk ratios and the chances of finding a significant association becomes far more likely. Findings can be made for those with heavy smoking husbands, light smoking husbands and ex-smoking husbands. Results can be divided by age, occupation, diet and social status. They can be split according to the type of lung cancer the cases are suffering from (there are four), as well as other cancers, heart disease, stroke and overall mortality. They can be rearranged according to the type of exposure (childhood, adulthood, spousal, mother, father, sibling, social, workplace) and, finally, the risks can be adjusted as the author sees fit in order to account for confounding variables.

There is a natural tendency for epidemiologists to want to show a positive result if only because null studies are of little interest and are less likely to be published. This tendency is particularly strong when the issue relates to secondhand smoke and when the researcher has a personal bias. From the very outset, there was a hope and expectation that passive smoking was indeed linked to lung cancer in nonsmokers. This prevailing bias has led to studies being written up in such a way that emphasised the results that supported the passive smoking theory and ignored the vast majority that did not.

How these results are presented is entirely down to the authors and their interpretation invariably moulds the report's summary and the accompanying press release. They may choose to publish only the results that appear to support the hypothesis or, if they tabulate the rest of the findings, write up their paper in such a way as to stress positive associations and downplay the null findings. If 99 results support the null hypothesis and one supports the a priori hypothesis, it is the single positive association that makes the headlines.

What follows is every peer-reviewed study of nonsmoking wives ever published with the editorialising stripped away to reveal the data in its pure form. Doctoral theses and dissertations are not included unless they have subsequently been published in a book or scientific journal. When results have been published more than once (eg. Hirayama, Fontham), the most recent version has been reviewed. Where confounding factors have been accounted for, the adjusted odds ratios have been used.

The studies are listed in descending order of size, with the studies with the largest sample group listed first. The order of the studies is important since those with the largest sample group are likely to offer the most accurate results. The reader will notice that the higher relative risks appear towards the bottom, where the smallest and least reliable studies lie. If one examines the results from the ten largest studies it is very difficult to view them as anything other than a random assortment of numbers hovering either side of 1.0. In order, they appear: 1.29, 1.11, 0.70, 1.03, 1.53, 1.10, 0.89, 1.10, 0.90 and 0.96. Between them, they give an average relative risk of 1.06 which is so close to a zero risk that if it were not so political, the issue of passive smoking would have been quietly shelved years ago. The smaller studies lift the average slightly higher - as the EPA found to their benefit - but some of these involve just 8 or 9 women and, with apologies to their authors, they are meaningless.

The results of studies that have investigated childhood exposure, workplace exposure or the effect on men are no more consistent or compelling than those involving nonsmoking wives (and the reader is encouraged to seek them out) but there are fewer of them and so the studies listed below provide the best evidence regarding the passive smoking theory.

The number of subjects is based on the total number of female lung cancer cases involved in the study. The relative risks are taken from the tabulated evidence given in the original study in most cases. In a small minority of cases, relative risks are not shown and in these instances the risks have been calculated from the available data. Statistically nonsignificant risks that exceed 1.0 are marked "(null)" and those that fall below 1.0 are marked "(negative)". Statistically significant findings are marked with an asterisk.

Now view The Evidence