Statistics from Altmetric.com
Recently the results of a comprehensive epidemiological follow up study of cancer mortality in cohorts with occupational exposure to acrylamide was published.1 With the exception of a weak significance for a raised incidence of pancreatic cancer the study arrived by and large at the conclusion that there is “little evidence for a causal relation between exposure to acrylamide and mortality from any cancer sites”. The study updates and confirms an investigation 10 years earlier of the same cohorts.2 The analysis was based on standardised mortality ratios (SMRs) in comparison with United States national or relevant county mortality statistics. It exemplifies the shortcomings of epidemiological studies of this kind to detect moderate influences of specific causative factors on cancer mortality or incidence. The investigators state that they have carried out “the most definitive study of the human carcinogenic potential of exposure to acrylamide conducted to date”. The results, however, pose questions. Could unacceptable risks be detected? Which risks would have been expected?
For the workers in the United States the average cumulative exposure is given as 0.25 mg/m3.y. (We assume this to correspond to exposure of the whole factory staff to 0.25 mg/m3 for 365 8 hour working days). At an alveolar ventilation rate of 0.2 l/kg.min this exposure would mean a cumulative uptake of about 9 mg acrylamide per kg body weight. This dose corresponds to a lifetime (70 years) uptake of 0.35 μg/kg.d. According to the estimate of the United States Environmental Protection Agency3 this would correspond to a cancer risk of 1.6×10-3. An estimate based on the multiplicative model4 would arrive at roughly a 3 times higher risk, 5×10-3. With a cancer mortality in the western world countries of 0.18, these figures correspond to a 1%–3% increase of the cancer mortality risk (RR)—that is, an RR of 1.01–1.03. As about one fifth of the workers were defined as exposed (at⩾10–3 mg/m3.y) the relative risk in the exposed group due to inhalation of acrylamide may have been about 1.05–1.15.
Although it is doubtful that these risk increments could be considered negligible, they would not be detectable in a study of the present kind. As uptake through the skin often occurs in addition to inhalation of acrylamide it is possible that the true risk increments are considerably higher. If we assume the total relative risk (from inhalation plus dermal uptake) to be in the range of 1.1–1.2, it is a pertinent question whether this risk increment is detectable within the large body of material studied by Marsh et al.1
Like many other materials of similar kinds the data are far from ideal for epidemiological analyses. The main reasons for this are the skewed distribution of duration of employment, the incompleteness of data for smoking, and the healthy worker effect. The healthy worker effect leads to a deficit in death rates from all causes, in the present study by about 20% for all causes except cancers. Deficits in SMR for all malignant neoplasms and for certain tumour types are also often significant, although with a disturbing influence of a significantly increased SMR for lung cancer in an earlier period. (The significant decrease in deaths from lung cancer as well as deaths from diseases of the circulatory system from 1925–83 to 1984–94 would be compatible with a drastic reduction in smoking, before 1984.) It is expected that the healthy worker effect comprises cancer, at least to some extent, as well as other causes of death.
A straightforward way of overcoming the healthy worker effect is a within cohort analysis of the regression of mortalities or incidences on the estimated dose. Marsh et al 1 have done this for each of a few selected tumour sites. Due to too few observed deaths in each dose interval the statistical power of this material is, however, too small to show anything.
This analysis of individual sites, avoiding a pooling of data that would increase the statistical power, illustrates the widespread dogma that different cancer types are affected specifically by carcinogens. It has been shown for a few mutagenic carcinogens including acrylamide that a linear multiplicative model, Pj=P0 j (1+β D), can be fitted to experimental cancer incidence data and, for radiation, to human data.5 Pj and P0 j are the total and background risks of tumour at site j, D the dose and β a relative risk coefficient that is (at least approximately) the same for all tumour sites j. β is thus applicable to pooled data for groups of sites or for all (responding) sites. Although analysis of death risks associated with specific tumours has its indisputable value, a restriction of estimation of significance to individual sites leads as a main effect to a loss of statistical power. For related reasons the identification of certain sites as “interesting”, with reference to response to acrylamide in animal experiments, is mostly a consequence of the pattern of background incidences P0 j in the animal strain used.
The authors of the paper1 possess information of extreme value in further efforts to clarify the carcinogenic potency of acrylamide. In view of the importance of this question we urge the authors of the paper to continue their work, particularly with analyses of regression on pooled data, primarily for all cancers, with and without exclusion of sites related to smoking.
Marsh et al reply
Granath et al take issue with our update of a cohort of acrylamide workers from three United States plants1-1 claiming that “it exemplifies the shortcomings of studies of this type to detect moderate influences of specific causative factors on cancer mortality or incidence.” To support their contention that we overlooked a small but “unacceptable” increase in cancer risk, they performed a crude quantitative risk assessment. Granath et al suggested that we perform a within cohort dose-response analysis with all malignant neoplasms as the end point as a means of attaining greater statistical power. They further contend that initial focus on specific cancer sites implicated in previous experimental animal studies is mostly a consequence of the pattern of background incidences in the animal strain used. Although choosing a generic health outcome such as all cancer sites combined will certainly increase statistical power, it also greatly reduces the ability to evaluate the all important specificity of an exposure-response relation. It is unlikely that even the most potent carcinogenic agent will increase the risks of all cancer sites to a level that can be detected with epidemiological methods.
We were fully justified in using cancer site specific findings as the focus of our epidemiological investigation. The use of cancer site specific findings from experimental animal studies to formulate initial testable aetiological hypotheses for human studies is an effective, accepted method commonly used in occupational epidemiological research. Animal studies can be particularly helpful when investigators are faced with a paucity of extant epidemiological evidence such as in the case of acrylamide. This practice does not preclude, however, the exploratory investigation of other non-implicated sites as long as the related findings are interpreted in the light of their hypothesis generating nature.
We agree that for many of the initial cancer sites examined in our study, the statistical power to detect a moderate excess in mortality (1.5 to twofold or greater) was low, a point considered in the discussion section of our paper. However, the power of our study to detect a twofold or greater excess in lung cancer, the end point of primary concern, at the one sided 5% significance level was in the excellent range (0.87), as would be the power to detect a similar excess of pancreatic cancer in a future update of this cohort.
Granath et al overlook a fundamental point—occupational cohort studies of the type we used to evaluate cancer mortality risks among workers exposed to acrylamide are neither designed nor necessarily well suited for quantitative risk assessment. Occupational cohort studies are purposely not designed to detect small excesses in the range of 5%–15% deemed by Granathet al as unacceptable. The primary reason for this is that excesses of this magnitude could easily be due, at least in part, to one or more confounding factors. Observational epidemiological studies usually cannot discriminate among such small mixed effects, and are generally most useful for detecting increases in risk that exceed 50%–100% as these are unlikely to be due to uncontrolled confounding. Considerations of statistical power notwithstanding, the fact remains that our study is the largest and most comprehensive study of exposure to acrylamide conducted to date, and will continue to provide useful epidemiological information through future updates and analysis.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.