Evolution of cancer frequency over the last 20 years

Between 1980 and 2000, the annual number of new cancer cases in France has gone up from 170,000 to 278,000 (63%), the rise being slightly more pronounced in men (97,000 to 161,000; 66%) than women (73,000 to 117,000; 60%) (Table 1.I).

Table 1.I

Increase in the number of cancer cases in France.

It is estimated that 45% of the rise in the number of cases are due to French society's demographic evolution (population growth and ageing) and 55% to an increase in the frequency (incidence) of cancers for each age group (Figure 1.1).

Figure 1.1

The rise in incidence varies according to cancer localization. The publication by Remontet et al. (2002) gives the mean annual evolution rate (%) of the age-standardized incidence of each cancer localization between 1978 and 2000. The following bar graphs (Figures 1.2 and 1.3) show the cancer localizations for each sex with a rise in annual incidence rate greater than or equal to 1% by order of decreasing incidence rise.

Figure 1.2

Figure 1.3

In men, the five cancers that show the most marked incidence increase between 1978 and 2000 are skin melanoma, prostate cancer, liver cancer, pleural mesothelioma, and malignant non-Hodgkin's lymphoma. In women, they are pleural mesothelioma, thyroid cancer, lung cancer, skin melanoma and renal cancer.

In terms of public health, estimating the relative weight of a given cancer localization requires that one also consider the mean annual incidence level. Thus, for example, a 10% annual increase in mesothelioma incidence would lead to the appearance of about 100 new cases each year while a three-fold lower increase in lung cancer would lead to the appearance of 1,000 new cancer cases.

Aside from the criterion of annual variation percentage, it is interesting to consider a criterion that combines mean frequency and level of increase. Table 1.II shows a ranking of the 15 cancers with the highest mean annual number of new cases (with a threshold of 30) in France from 1980 to 2000.

Table 1.II

Cancers with the highest mean annual number of new cases in France between 1978 and 2000.

Judging from the mean annual number of new cases due to incidence increase, the two predominant localizations are prostate and breast cancer, which alone represent over 50% of the new cancer cases due to a rise in incidence (Figure 1.4). While exhibiting one of the highest relative increases, mesothelioma only ranks 14 in order of importance by this particular criterion.

Figure 1.4

The choice of criterion to classify cancers as a function of their evolution over the last 20 years depends entirely on the nature of the question to be addressed. Should one wish to focus on the quantitative aspect, the criterion that combines mean incidence level and relative annual variation seems to be the most appropriate.

1. Identification of cancers exhibiting an incidence and mortality increase

Table 1.III shows cancers exhibiting an incidence increase that contributes very significantly to the rise in total cancer (in decreasing order).

Table 1.III

Increase in cancer number, incidence and mortality between 1980 and 2000 (from Remontet et al., 2002).

The table also shows whether these cancers are not only on the increase but also exhibit a concomitant increase in mortality. A rise in incidence without a parallel increase in mortality can be explained at least in part by modifications in health care practice (screening and diagnosis). Indeed, one might suppose that incidence increase is essentially due to cancers with better prognosis particularly as a result of improved screening procedures and earlier diagnosis. A rise in incidence without an increase in mortality could also be explained by substantial therapeutic progress for particular cancer localizations.

Finally, one might suppose that a joint increase in incidence and mortality reflects a risk increase linked to the fact that the prevalence of risk factors is on the rise or rose over a time period compatible with the latency period of the cancers involved. Such factors may be linked with individual behavior (sedentary lifestyle, alcohol, active smoking, sun exposure) and environmental factors (occupational exposure, passive smoking, and atmospheric, earth or water pollution).

Table 1.IV presents cancers exhibiting a parallel increase in incidence and mortality in France, in decreasing order of incidence.

Table 1.IV

Cancers exhibiting a parallel increase in incidence and mortality in France between 1980 and 2000, in decreasing order of incidence.

In conclusion, among cancers that show a rise in incidence, only non-Hodgkin malignant lymphomas, lung cancers, liver cancers, brain and CNS cancers, pancreatic cancers and pleural mesothelioma have presented a parallel rise in mortality over the last 20 years.

The incidence increase for breast and prostate cancers represents more than 50% of the annual rise in cancers and do not exhibit a parallel evolution of incidence and mortality. Not surprisingly, both are cancers for which there has been a progressive development of screening over the last 20 years.

References

1. Remontet L, Buemi M, Velten M, Jougla E, Esteve J. Évolution de l'incidence et de la mortalité par cancer en France de 1978 à 2000. Éditions InVS, 2002: 217p.

Cancer and causes of premature deaths in the European Union

Deaths by cancer represent a quarter of the total number of deaths that occur annually in the fifteen EU member countries (960,000 deaths by cancer per year). In the case of premature deaths (under 65), the proportion is as high as 37% (261,000 deaths) with cancer representing the first cause of death, ahead of cardiovascular mortality. For men (534,000 deaths by cancer), lung cancer represents 28% of total cancer mortality, followed by cancers of the intestine (11%) and prostate (10%). For women, breast cancer is predominant (1 death of 5) ahead of cancer of the intestine (12%) and lung cancer (11%).

For most cancer localizations, a general improvement in mortality levels in the EU from 1994 to 1999 has been observed (Tables 2.I and Tables 2.II). The most significant progress is seen for cancers of the stomach (both sexes), of the bladder and kidney (men), and of the uterus. Conversely, the frequency of occurrence of certain localizations is unchanged: pancreas, liver, lymphomas and leukemia (both sexes), and upper respiratory tract (women). European death levels have risen for two cancer localizations: lung cancer in women and skin melanoma in men. The decrease in death rate for men under 65 has been slightly more pronounced, irrespective of anatomic site, than for the population in general. In the case of women, the reduction in death rate is of a similar order for both premature deaths and total deaths. There is, however, a strong progression of female mortality under 65 by cancer of the lung and upper aero-digestive tract (UADT).

Table 2.I

Death rate by cancer in the European Union and in France,¹ male.

Table 2.II

Death rate by cancer in the European Union and in France,¹ female.

Death rate by cancer in France relative to the European Union

A strong disparity in levels of mortality by cancer can be observed overall and according to cancer type between countries of the European Union (Table 2.III).

Table 2.III

Standardized death rates by cancer in EU countries,¹ male; all ages.

Cancer mortality in men

France has the highest male death rate for cancer, followed by Belgium and the Netherlands. The lowest death rates are found in Sweden, Finland and Greece. Thus for instance, the death rate is 50% and 20% higher in France than it is in Sweden and the United Kingdom respectively.

The unfavorable situation for men living in France is for the most part explained by the weight of UADT, liver, and lung cancers. Thus, 65% of the total overmortality by cancer in France compared to the United Kingdom is due to a higher death rate for the three above localizations. For UADT cancers, France ranks first among EU countries with a major variation from all other countries. Regarding liver cancer, Italy, France and Greece take the lead with death rates of a similar order in all three countries. However, while mortality is decreasing markedly in Italy and Greece, French mortality rates are stagnating. With respect to lung cancer, Belgian and Dutch death rates are the highest, but France ranks in the European average. However, although a drop in the number of lung cancers can be observed in many countries, French death rates are not falling significantly. In the case of mortality under 65, France is among the countries with the highest mortality for lung cancer, along with Belgium and Spain.

Apart from the above specificities, France's situation is average or favorable for other types of cancers in men, while particularly high death rates are observed for certain localizations in other countries: stomach cancer in Portugal, prostate cancer in Sweden, colorectal cancer in Ireland and Denmark, and skin melanoma in Denmark and Sweden.

In the under 65 population (Table 2.IV), France has the highest male mortality rate (together with Spain and Belgium). This is for the most part explained by the high frequency of lung cancers in all three countries. Similarly, France ranks at the top of all EU countries for "premature" mortality by UADT cancer (with Luxemburg) and by liver cancer (with Italy).

Table 2.IV

Standardized death rates by cancer in EU countries,¹ male; under 65.

Cancer mortality in women

Unlike men, French women occupy a generally favorable position within the EU (Table 2.V and 2.VI). For all localizations, female death rates by cancer are the highest in Denmark, Ireland, the United Kingdom, and the Netherlands. The highest mortality rates are observed in Denmark and the Netherlands for breast cancer; in Denmark, the United Kingdom, and Ireland for lung and UADT cancers; in Denmark and Germany for colorectal cancer; in Portugal for stomach cancer; and in Greece and Italy for liver cancer. In Denmark, lung cancer constitutes today the most frequent localization in women, ahead of breast cancer. In the United Kingdom, these two types of cancer exhibit the same frequency among the female population.

Table 2.V

Standardized death rates by cancer in EU countries¹; female; all ages.

Table 2.VI

Standardized death rates by cancer in EU countries1; female; under 65.

The evolution of female death rates by cancer between 1994 and 1999 shows a downward trend (with the greatest decrease in Austria and Luxemburg). Compared with this general trend, however, stagnation in the mortality levels is seen for certain countries, including France. The stability of death rates in France is for the most part explained by the observed trends with lung, UADT, breast, and uterine cancers. Women living in France currently exhibit the highest increase in lung cancers (along with Luxemburg and the Netherlands), and this unfavorable position is even more marked for deaths under 65 (40% death rate progression between 1994 and 1999). For UADT cancers, Luxemburg and Austria exhibit the most marked rise in death rates (female rates are also rising in France but not as markedly as in these two countries). While death rates by breast and uterine cancers are falling in most EU countries, they are stagnant in France.

Variation between male and female death rates

Male overmortality by cancer is observed in all countries (Table 2.VII), but the largest variation between the sexes is seen in Spain and France, where the death rate is 2.2 times higher in men. Conversely, the differences between the sexes are the lowest in Denmark, Sweden and the United Kingdom (1.3). In terms of localization, male overmortality is the highest for UADT and lung cancers (4.5 for the whole of the EU). Male mortality rates by lung cancer are 12 and 7 times higher than the rates for women in Spain and France, respectively. Similarly, the figure for male overmortality by UADT cancer is about 10 in both countries. Conversely, the mortality ratios between sexes for these localizations are particularly low in Denmark, Sweden and the United Kingdom (about 2). Between 1994 and 1999, the level of male overmortality has remained stable, except in the case of lung and UADT cancers, where the difference between men and women has been reduced.

Table 2.VII

Men/women ratios of death rates by cancer in EU countries,¹ all ages; 1999.

Comparability of data from one country to another

It is necessary to investigate whether the comparisons between countries on which the preceding results are based are legitimate. The degree of comparability of data on causes of death in the EU is currently being analyzed as part of Eurostat's "Causes of death" Task Force (specific certification and codification practices by country). In this respect, a report is available that contains recommendations validated by all EU countries for improving the quality and comparability of data as well as a bibliography of the main scientific studies published on the subject (Jougla et al., 2001).

Although methods to obtain data on causes of death are being increasingly standardized over time, numerous analyses have highlighted the disparity between different countries' practices, whether in terms of medical certification of the causes of death or codification (selection of an initial cause for each death). Regarding medical certification, it is worth noting that all EU countries now use a death certificate similar to the one recommended by WHO in the International Classification of Diseases (ICD), which involves a common methodology to describe the morbidity process that lead to death (from an initial to an immediate cause). Similarly, EU countries apply ICD rules with increasing homogeneity to select the initial cause (on which death statistics are based) from the list of possible causes on the death certificate (WHO, 1992). The current trend toward automatic codification systems integrating identical decision rulings for choosing the initial cause of death will strongly contribute to achieve good standardization of the codification step.

Among different causes of death, cancer is one of the causes for which the degree of international comparability is most reliable (compared to pathologies such as cardiovascular diseases or violent deaths), especially when using fairly broad localization subgroups (Jougla et al., 2003). Nevertheless, certain localizations are a greater source of recording problems than others. Lung cancer, the most frequent tumor site for men, is characterized by a satisfactory concordance between information that comes from mortality and morbidity. In the case of breast cancer, studies based on a comparison of the "official" initial cause of death and the cause determined from other clinical sources conclude there is a slight underestimation in official statistics. For other anatomical sites, differences in recording practices may lead to comparability bias. These differences may be due to a difficulty in confirming the malignant or primitive character of a tumor (liver), or distinguish the primary site among neighbor organs (stomach-esophagus, pancreas-biliary tract, cervix-uterus), especially when clinical manifestations or histological typing are similar. Finally, the role of certain cancers with rather favorable prognosis in the death process can be overestimated compared to heavy associated pathologies (prostate and colon). Aside from potential bias linked to diagnostic difficulties or lack of precision in the declaration on the part of certifying practitioners, the data can be affected by random fluctuations, especially when death rates analyzed for a given country are low (skin melanomas, "premature" cancers of the urinary tract, etc.).

In conclusion, the results of analyses of mortality levels by cancer show major spatial disparities between EU countries. Thus, France's situation with respect to male cancer levels is not favorable. This unfortunate position is largely explained by very elevated death rates for lung, UADT, and liver cancers. While occupational exposure certainly carries a good deal of weight as a determinant of mortality levels (although this is not easily measured for lack of available data), we know that these types of cancers are strongly linked to two risk factors: alcohol abuse and smoking. Similarly, although France appears to be in a rather favorable position for female cancers, a very worrying trend is observed with respect to lung cancer. Given the delayed impact of tobacco consumption, the evolution of mortality is the consequence of the surge in women's smoking habits since the 1960s. The situation is all the more worrying that current indicators of female tobacco consumption are not encouraging (Beck and Legleye, 2003). Analysis of disparities by cancer category in the EU clearly shows the particularly negative and strong impact of alcohol and tobacco consumption on mortality levels in France compared to other countries. The repercussions of tobacco and alcohol consumption can also be observed in other countries, particularly in the case of women. With respect to breast and uterine cancers, no change has been recorded in France while many other countries exhibit a decrease in such cancers. In this respect, one might question the impact of national screening policies (Uhry et al., 2003), but the results must be confronted with incidence and survival data. This study also showed higher mortality rates for specific localizations (stomach, prostate, colon-rectum, skin melanomas) in other countries. Eurostat has established an Atlas that maps such causes of deaths down to the regional level and thus allows a more precise characterization of spatial mortality disparities across the EU (Jougla et al., 2003).

References

1. Beck F , Legleye S . Tabac à l'adolescence. Résultats de l'enquête Escapad BEH 2003 22–23 101 102 .
2. Boyle P , Ferlay J . Cancer incidence and mortality in Europe, 2004 Ann Oncol 2005 16 481 488 . [PubMed: 15718248]
3. Jougla E , Pavillon G , Rossollin F , De Smedt M , Bonte J . Improvement of the quality and comparability of causes of death statistics inside the European community Rev Epidemiol Sante Publique 1998 46 447 456 . [PubMed: 9950045]
4. Jougla E, Pavillon G. International comparability of causes of death data. In: Morbidity and mortality data--problems of comparability. Wunsh G, Hiancioglu A. (eds), Hacettepe University, Institute of Populations Studies, Hacettepe 1997: 75–95.
5. Jougla E, Rossollin F, Niyonsenga A, Chappert JL, Johansson LA, Pavillon G. Comparability and quality improvement in European causes of death statistics. Eurostat, Project 96/S 99–5761/EN, 2001: 190p. http://europa​.eu.int​/comm/health/ph/programmes​/monitor/fp_monitoring​_1998_frep_04_en.pdf.
6. Jougla E, Salem G, Gancel S, Michel V, Kurzinger ML, et al. Atlas on mortality. European Commission, Eurostat, Health statistics, 2003: 117p.
7. Jougla E , Salem G , Rican S , Pavillon G , Lefeve H . Disparités de mortalité par cancer dans l'Union européenne BEH 2003 41–42 198 201 .
8. Uhry Z , Fourme E , Jougla E , Cherie-Challine L , Ancelle-Park R , et les coordinateurs des structures de gestion départementales. Mortalité par cancer du sein dans les départements ayant mis en place depuis 1990 un programme de dépistage organisé du cancer du sein BEH 2003 4 19 21 .
9. World Health Organization. International classification of diseases, 10th revision. Geneva 1992.

Incidence of child cancers in France

Child cancer recording in France (excluding overseas territories) was accomplished in two major steps. Between 1983 and 1990, pediatric cancer registers were set up on a regional basis, first in Lorraine (1983), then in the PACA (Provence, Alpes, Côte d'Azur) region (1984), in Auvergne (1986), in the Rhône-Alpes region (1987), in Brittany (1991), and in the Limousin (1994). A register was also set up in the Val de Marne between 1990 and 2000. Incidence data from these registers over the period 1990–1999 is summarized in Table 3.I and has been published recently (Desandes et al., 1994).

Table 3.I

Incidence of child cancer in France, pediatric register data over the period 1990–1999 (from Desandes et al., 2004).

A second step led to national recording of child cancers (as of 1990 for malignant hemopathies and since 2000 for other tumors) while keeping regional registers active because they are more suited for work on follow-up and management of child cancers. The National Register for Child Cancers ( - RNLE) is directed by Jacqueline Clavel from Inserm Unit 170 in Villejuif. Registre National des Leucémies de l'Enfant (RNTSE; the national child solid tumors register) in Nancy is directed by Brigitte Lacour, who is also in charge of the Lorraine child cancer register. The recently published incidence data from RNLE (Clavel et al., 2004) are of course more precise with respect to malignant hemopathies than the regional data, but estimates are analogous to those made by regional registers. Apart from estimating cancer incidence and working on surveillance, the two registers also carry out etiological research as part of the INSERM Unit 170 research project on environmental and genetic risk factors for child cancer.⁵

Geographical variations of incidence

When focusing on histological typing, a particular localization, or age group, numbers become restricted and incidence estimates fluctuate, particularly when they are based on data from regional or departmental registers. This can make geographical and temporal comparisons difficult. Nonetheless, a number of facts have been well established:

• Burkitt's lymphomas are more frequent in countries with endemic malaria.
• Leukemia incidence is higher in industrialized than developing countries.

The incidence of leukemia is also higher in Hong-Kong than other Asian countries, in white compared to black people in the United States, and in non-Maoris compared to Maoris in New Zealand (Parkin et al., 1998). Socio-economic conditions and hygiene, nutritional factors, tobacco consumption, and accompanying environmental exposure certainly play a major role in accounting for this imbalance. Making abstraction of such intra-country variations however, the figures for Western countries are on the whole quite homogeneous. Within European countries (Automated Childhood Cancer Information System or ACCIS), fluctuations can be quite pronounced (Table 3.II), but no spatial organization or gradient can be identified by making international comparisons.

Table 3.II

Age-standardized incidence rates (ASR) of cancers in children under 15 in Europe (ACCIS, 2003).

Temporal variation in child cancer incidence

The trend toward an increase in cancer incidence is far from being claimed unanimously. In France (Figures 3.1 and 3.2), the trend over the period 1990–1999 is not one of increase. The data from the Lorraine child cancer register (Registre Lorrain des Cancers de l'Enfant), which started keeping records in 1983, does not show evidence of an incidence increase either.

Figure 3.1

Evolution in the incidence of child malignant hemopathies in France over the period 1990–1999 - Registre National des Leucémies et des Lymphomes de l'Enfant (from Clavel et al., 2004). ALL, acute lymphocytic leukemia; AML, acute myeloid (more...)

Figure 3.2

International data that goes further back in time shows a coherent increase in solid tumors from one country to the next during the 1980s. It is now well established that the break in the incidence curve was due to the brain imagery revolution (Smith et al., 2000). For other tumors, the data are contradictory and it is hard to separate improved diagnosis from better codification (to ensure quality recording) to explain the increase in the number of cases. In Western countries, no spatial organization has been identified that can distinguish countries with an incidence increase from those where incidence has remained stable. The recently published analysis of European data by ACCIS (Automated Childhood Cancer Information) (Steliarova-Foucher et al., 2004) shows an increase in the incidence of child cancers (0–19 years) from 1970 to 1990 in all age groups studied. The mean annual variation was 1% during the 1970s and 1980s, and 1.3% during the 1980s and 1990s. The increase is more significant in Eastern compared to Western European countries. However, the heterogeneity of registers included in this analysis urges one to interpret such results with caution.

Geographical and temporal incidence variations on a small geographical scale

Over the last 20 years, cancers, and in particular child leukemia, have been studied on a small temporal and geographical scale, when case occurrence can be considered over short distances. The aim of such studies is to show possible trends of spatio-temporal case clusters suggesting the presence of carcinogenic sources or localized epidemics. This issue remains hotly debated and has lead to the development of important new methodological techniques.

In conclusion, the increased incidence in child cancers often reported in the media is based on inconsistent reports and cannot be accepted as an established fact. Conversely, the possible link between environmental exposure and the occurrence of several types of child cancers, particularly malignant hemopathies and brain tumors, is increasingly well documented. However, a clear causal relationship has only been shown in the case of exposure to high doses of ionizing radiation. The role of infectious agents in Burkitt's lymphoma and Hodgkin's disease (EBV) has also been clearly demonstrated and is strongly suggested in acute leukemias and, to a lesser degree, in brain tumors. Arguments are slowly accumulating to suggest that low doses of radiation (particularly of natural origin), very low frequency magnetic fields, pesticides, and atmospheric pollution generated by automobile traffic may also play a role. Whether it is responsible or not for a detectable increase in cancer incidence, the environment is probably involved in the occurrence of a number of child cancers.

References

1. Automated Childhood Cancer Information System (ACCIS). A system of provision, presentation and interpretation of data on cancer incidence and survival of children and adolescents in Europe. http://www-dep​.iarc.fr/accis.htm [accessed 3 September 2003].
2. Birch J, Boffetta P, Ahlbom A, Chybicka A, Clavel J, et al. Baseline report on childhood cancer in the framework of the European environment and health strategy. European Commisssion (DG ENV.). SCALE Working Group on Childhood Cancer. http://www​.brussels-conference​.org/project.htm [accessed 15 November 2004].
3. Clavel J , Goubin A , Auclerc MF , Auvrignon A , Waterkeyn C . et al. Incidence of childhood leukaemia and non-Hodgkin's lymphoma in France: National Registry of Childhood Leukaemia and Lymphoma, 1990–1999 Eur J Cancer Prev 2004 13 97 103 . [PubMed: 15100575]
4. Desandes E , Clavel J , Berger C , Bernard JL , Blouin P . et al. Cancer incidence among children in France, 1990–1999 Pediatr Blood Cancer 2004 43 749 757 . [PubMed: 15390289]
5. Parkin DM, Kramarova E, Draper GJ, Masuyer E, Michaelis J, et al. International incidence of childhood cancer. IARC Scientific Publications No 144 (II), Lyon, 1998.
6. Smith MA , Freidlin B , Ries LA , Simon R . Increased incidence rates but no space--time clustering of childhood astrocytoma in Sweden, 1973–1992: a population-based study of pediatric brain tumors Cancer 2000 88 1492 1493 . [PubMed: 10717635]
7. Steliarova-Foucher E , Stiller C , Kaatsch P , Berrino F , Coebergh JW . et al. Geographical patterns and time trends of cancer incidence and survival among children and adolescents in Europe since the 1970s (the ACCIS Project): an epidemiological study Lancet 2004 364 2097 2105 . [PubMed: 15589307]

Hill's Criteria

In spite of these theoretical restrictions, epidemiologists endeavor to judge the degree of causality plausibility of a relationship on the basis of available data, both to improve knowledge about disease determinants and to propose preventive measures. This approach is complicated by the fact that epidemiological studies are more often than not observational (cohort and case-control studies, for example) rather than experimental. In this respect, it must be pointed out, along with Rothman and Greenland (1998), that results from experimental studies (randomized trials, for instance) can show discrepancies between studies or be interpreted in diverging ways and that conversely, the non-experimental nature of a discipline does not necessarily impede major scientific progress (e.g., understanding the movement of planets, plate tectonics, or the evolution of species).

In practice, the chosen approach is based on a set of criteria to be confronted to judge the degree of causality plausibility of a relationship. Such criteria are examined one by one and then analyzed synthetically to progress and generally assess plausibility. It is worth mentioning that such an overall judgment does not entail assigning a score value or obtaining a numerical result even if categories of causality plausibility levels are used by certain organizations (see below).

The most commonly used criteria are those proposed by Hill (1965). These criteria include and add to the criteria proposed in 1964 in the American Surgeon General's report about the effects of smoking on health (United States Department of Health, Education and Welfare, 1964) and are akin to the rules proposed earlier by Hume (1739) and to Mill's inductive canons (1862). There are nine such criteria (Table 4.I), the first five of which characterize the nature of the relationship between the exposure factor under consideration and disease risk as a function of the results obtained from epidemiological studies. The last four criteria serve to put the results of the epidemiological studies into perspective relative to biological knowledge pertaining to the relationship under study. It is worth noting that some versions of this set of criteria only take into account seven criteria by doing away with the last two of the second group of criteria, or even six criteria by fusing the first two criteria from the second group.

Table 4.I

Hill's causality criteria (1965) .

Criteria that characterize the nature of the relationship

Among the five criteria that characterize the nature of the relationship, the first four are involved with the results of individual epidemiological studies. They can therefore be verified by some studies and not others. They may also be involved in synthesis of epidemiological studies in the form of meta-analyses. The fifth criterion (reproducibility) concerns the confrontation of results from various epidemiological studies.

Contextual criteria

The last four criteria involve biological knowledge about the association under scrutiny.

Systems of classification of causal relationship plausibility

To propose concrete prevention measures in spite of the difficulties outlined above, systems to classify the degree of plausibility of possibly causal relationships have been proposed and put into practice. Such systems integrate elements akin to Hill's criteria. The two best known elements are cancer-related and emanate from the International Agency for Research on Cancer (IARC), which is affiliated with WHO (http://www-cie.iarc.fr/), and the United States Environmental Protection Agency (U.S. EPA) (1999).

For example, IARC experts regularly meet to pass judgment on the carcinogenic or non-carcinogenic nature of exposure to all kinds of substances present in the environment and publish these evaluations in the form of monographs. Both human and animal data are taken into account to agree on a general evaluation for each substance and assign it to a group in the five-level classification of plausibility that a substance is carcinogenic in humans:

• Group 1: carcinogenic substance
• Group 2A: probably carcinogenic substance
• Group 2B: possibly carcinogenic substance
• Group 3: unclassifiable substance
• Group 4: probably non-carcinogenic substance

In the first 80 monographs that IARC has published (1972–2002), it has evaluated 878 substances and has classified 87, 63, 234, and 493 of these in groups 1, 2A, 2B, and 3, respectively, and one in group 4. Although these evaluations are of great value, it must be remembered that the elements taken into account are similar to Hill's criteria and are therefore subject to similar limitations.

In conclusion, establishing a causal relationship is useful both to further knowledge and propose public health prevention policies. It is, however, a very complex task. Hill's causality criteria are helpful in evaluating the causal nature of an association in spite of their many limitations and the impossibility of reaching a formal conclusion. Systems of exposure classification according to the degree of plausibility of a causal association have been formulated. They comprise elements similar to Hill's criteria and are commonly used.

References

1. HILL AB . The Environment and Disease: Association or Causatio? Proceed R Soc Med 1965 58 295 300 . [PMC free article: PMC1898525] [PubMed: 14283879]
2. HUME D. A Treatise of Human Nature. Oxford University Press, Oxford, 1888, Second Edition, 1978 (Original publication, 1739)
3. INTERNATIONAL AGENCY FOR RESEARCH ON CANCER (IARC) WORKING GROUP FOR THE EVALUATION OF CARCINOGENIC RISKS TO HUMANS. Tobacco Smoke and Involuntary Smoking. IARC Monographs on the Evaluation of Carcinogenic Risks to Humans, Lyon, France, 2004, Volume 83. [PMC free article: PMC4781536] [PubMed: 15285078]
4. KLEINBAUM DG, KUPPER LL, MORGENSTERN H. Epidemiologic Research : Principles and Quantitative Methods. Chap 2, John Wiley and Sons, New York, 1982 : 529p.
5. LANES SF , POOLE C . Truth in Packaging? The Unwrapping of Epidemiologic Research J Occup Med 1984 26 571 574 . [PubMed: 6481496]
6. Mill JS. A System of Logic, Ratiocinative and Inductive. Chap 2, Fifth edition, Parker, Son and Bowin, London, 1862. In: Preventive and Community Medicine. Clark DW, Macmahon B. (eds), Little Brown, Boston, 1981, 2^eédition.
7. Popper KR. The Logic of Scientific Discovery. Harper and Row, New York, 1968.
8. Rothman KJ, Greenland S. Modern Epidemiology. Lippincott-Raven, Philadelphia, 1998, 2nd Edition, Chap 2.
9. United States Department of Health, Education and Welfare. Smoking and Health: Report of the Advisory Committee to the Surgeon General of the Public Health Service. Government Printing Office, Washington, DC, PHS Publication No. 1103, 1964.
10. United States Environmental Protection Agency. Guidelines for Carcinogenic Risk Assessment. Risk Assessment Forum. NCEA-F-0644. US-EPA, Washington, DC, 1999.

Respective contributions of environment and heredity

The relative contribution of environmental and genetic factors in the appearance of cancer is not easy to determine. Studies carried out on thousands of Scandinavian twins have nevertheless permitted an appreciation of the relative contribution of the environment versus heredity in the genesis of different cancer types (Lichtenstein et al., 2000). The results of these studies have been confirmed in a study involving millions of individuals in Sweden and designed to evaluate the familial nature of many cancers (Czene et al., 2002). Other studies carried out on couples have also improved our knowledge on the origins of cancers and, in particular, the important contribution of lifestyle (Hemminki et al., 2001).

The view that a clear distinction needs to be made between genetic mechanisms and environmental factors now appears far too narrow. Finally, various environmental components can potentially interact: thus in third-world countries, the appearance of hepatic cancers is amplified by Hepatitis B viral infection, aflatoxin contamination, and the genetic profile that determines the metabolism of this substance.

In summary, the origins of cancers must be examined according to cancer type and cancer localization while taking into account the interactions between genetic factors and the various components of our environment.

Various routes of contamination

The study of contamination via the environment must take several factors into account. Contamination can occur by ingestion, inhalation, or transdermal absorption. In the case of chemicals, assessing true toxicity requires good knowledge of their distribution in the environment, routes of contamination, as well as kinetic and dynamic properties in the organism.

The only pollutants discussed in the following presentation are chemicals. However, experts should also eventually study other environmental factors in a future collective expertise.

Classification of chemical pollutants

Chemical pollutants can either be classified according to structure or their probable mode of action. Both classifications present certain advantages.

Classification according to chemical structure is as follows:

• Aromatic polycyclic hydrocarbons (benzopyrene)
• Organochlorides and organobromides (pesticides, dioxin, PCB, polybromides)
• Aromatic amines
• Organophosphorus compounds (sarin, chlorpyrifos)
• Nitrosamines
• Fibers: asbestos
• Heavy metals
• Other compounds (toxins such as aflatoxin)
• Mixtures: tobacco, fine particles, tars

Classification by major mode of action is as follows:

• Direct genotoxicants: physical agents, benzopyrene, aflatoxin
• Non-genotoxicants: substances with their own cell signaling apparatus (dioxin-AhR receptor, pesticides-PXR receptor); endocrine disturbing substances (activation or inhibition of cell signaling, estrogen mimetic substances, organochloride pesticides); enzyme disturbing substances (organophosphorus compounds); cell stressors (oxidative stress, asbestos, metals, dioxin)
• Indirect genotoxicants leading to cumulative toxicity (particles, asbestos, inflammation) or multiple toxicities (mixtures or non-mixtures)

The two classifications are of course not independent of each other, but agents that are different in nature can clearly be genotoxic (radiation and xenobiotics, for example) while a particular type of agent such as organochloride pesticides can have various mechanisms of action.

Recent progress in the study of the mode of action of chemical pollutants

We are mainly interested here in advances in the field of non-genotoxic pollutants. Such pollutants activate different types of receptors that can be classified into two major categories: xenobiotics receptors sensu strictu (AhR receptors for dioxin and aromatic polycyclic hydrocarbons; PXR receptors capable of linking drugs with pesticides; the CAR receptor, the role of which remains to be elucidated with respect to environmental pollutants). The main function of such receptors is to allow the organism to adapt to xenobiotic influx because they are responsible for the induction of enzyme systems involved in elimination of xenobiotic substances. The second set of relevant receptors in terms of the environment is a group of receptors for endogenous substances, such as hormonal receptors, but nonetheless susceptible to modulation by polluting compounds (e.g., organochloride pesticides and the estradiol receptor). This is a case of "illegitimate receptor activation" leading to endocrine or metabolic disturbance (Figure 5.1).

Figure 5.1

It is interesting to note that toxicity can arise from interaction of a pollutant with its "legitimate" receptor (for example dioxin and the AhR receptor) as well as receptors for endogenous substances. It is also worth noting that the above receptor classification is oversimplified in that xenobiotic receptors can also bind endogenous substances (PXR receptors bind biliary acids, for instance). Finally, the concept of nuclear receptor affinity must be taken into account: the estradiol receptor has a 1,000-fold greater affinity for the natural hormone than for organochloride pesticides. In general, xenobiotics have moderate affinities for their receptors (on the order of the μM), which means it is necessary to verify the relevance of the results obtained; dioxin is the only substance that presents a strong affinity (on the order of the nM) for its receptor (AhR).

There are numerous models to study pollutant toxicity. There are also many in vitro or ex vivo as well as animal tests available. These tests are well characterized for genotoxic compounds, even if the value of such tests is sometimes controversial. Studies on the mechanism of action of non-genotoxic substances are not so well structured and depend on the expertise of each laboratory as well as the availability and relevance of animal and cellular models. It is important to emphasize that the mechanism of action of genotoxic and non-genotoxic compounds is different, and that it is therefore hazardous to transpose knowledge relative to one of these categories to the other. This is particularly important with respect to mathematical models for predicting toxicity at low doses. Such models are often based on linearity of toxicity as a function of dose for genotoxic compounds, but this is very controversial in the case of other compounds.

Nevertheless, recent studies are not so intent on separating genotoxic from non-genotoxic compounds. Numerous compounds that are not genotoxic induce an oxidative stress that can alter DNA and thus lead to genotoxicity (dioxin). Other compounds have themselves a non-genotoxic action while their metabolites are capable of forming DNA adducts (benzopyrene). The classical model of chemical carcinogenesis in rodents (initiator versus promoter) may be oversimplified under such conditions. Furthermore, the study of genic alterations in human tumors shows that there is a succession of alterations leading ultimately to the appearance of cancers. Finally, apart from the effects on tumor initiation and promotion, pollutants might cause tumor progression and possibly facilitate cancer dissemination and metastases. The latter issue has unfortunately received little attention as a result of a lack of relevant experimental models.

What is the use of experimental toxicology?

The aim of studies in the field of health and the environment is to establish a relationship between an environmental factor and the appearance or the aggravation of cancer. Epidemiological studies are necessary and could in the first place focus on populations that are particularly exposed, whether it is for professional, accidental or geographical reasons. The difficulty is then to generalize risk findings to the general population, which is subject to weaker exposure.

To establish such relationships, a major difficulty is to identify and measure exposure. Indeed, exposure may have occurred many years previously, in which case the challenge is to obtain reliable data on the quantity of exposure. There may also be residual quantities of a chemical or a persisting biological effect. This is a difficult step that weakens a large number of studies.

Furthermore, epidemiological studies can suffer from biases or lack sufficient power. Reported toxicity can be attributable to compound mixtures and it is difficult to know which of the substances or set of substances in the mixture is actually responsible for the effect.

Thus it is necessary to establish the biological plausibility of a carcinogenic effect to define the true incurred risk. The contribution of toxicology is not equivalent for all studies, but it is unavoidable in situations where epidemiological criteria turn out not to be sufficiently powerful.

Toxicology and demonstration of mechanisms of toxicity also play an important role in the search for exposure markers, work on predicting toxic effects, and preventive action. In this respect, good knowledge of mechanisms of action also plays an economic role by helping prevent industry from spreading potentially dangerous substances.

To summarize, the experimental toxicological approach is useful:

• to establish the biological plausibility of the cause-effect relationship
• as a predictive approach
• as a public health preventive approach

It also presents an obvious economic advantage for industry.

Difficulties associated with experimental toxicology

Studies on the mechanisms of action of a pollutant or an environmental factor are carried out using model systems. Animal models are often used but their relevance to the human situation is sometimes controversial or poorly established. Ex vivo or in vitro models derived from human samples are also used but criticized precisely for not being whole organism studies. It is often necessary to have an array of arguments to confirm the relationship between the biological properties of an environmental factor and its role in the appearance of cancers.

One must nonetheless emphasize that knowledge on the mode of action enables one to know better whether observations made on animals can be transposed to humans. Examples of relative failure with animal models clearly illustrate this point (Meek et al., 2003).

For instance, the renal carcinogenicity of limoneme in male rats is linked to its binding to the alpha-2-microglobulin protein. Now, this protein is specifically found in rats and has no equivalent in man. This animal model is therefore not transposable.

Peroxisome proliferators are potent hepatic carcinogens in rodents but fibrates are not carcinogenic for man. This difference can be explained by a discrepancy in the number of receptors (PPAR receptors) for such compounds between humans and rodents. Furthermore, human and rodent PPAR receptors seem to have distinct, sometimes diametrically opposed, genic effects.

Dioxin is a potent carcinogen for certain rodent strains, whereas its effect in humans is relatively moderate. Now, in sensitive rodent strains, the affinity of the AhR dioxin receptor for this pollutant is 10 to 100 times higher than the affinity for the corresponding human receptor, which explains the observed difference in sensitivity.

Thus, better knowledge of the mode of action enables one to judge more competently the relevance of an animal or cellular model for testing the toxicity of a given pollutant in humans.

New methodologies

Experimental toxicology can certainly benefit from new technologies in genomics, proteomics, and metabonomics. Such large-scale approaches should help get a better understanding of mechanisms of action, compare effects in different species, provide new biological markers, and contribute to developing predictive toxicology. Other useful technologies are analytical methods for the dosage of pollutants as well as transgenic approaches aimed at "humanizing" animal models (e.g. mice models) to make them closer to the human situation and improve the possibilities of transposing results from animal to humans.

Questions and future challenges for experimental toxicology

In spite of recent progress, numerous issues of interest for other fields of activity in the area of health and environment remain unresolved. We ought to be able to understand and predict more efficiently the effects of chronic exposure to low doses of pollutants, which remain an unresolved problem for cancer in particular. We should have more insight into the effects of mixtures, because most contaminations are multiple in nature. Although we are beginning to understand the mode of action of certain pollutants, we still know very little about the effects of mixtures and about the interactions between modes of action (synergy, opposite effects or independence). This is crucial for contaminants that are often present in conjunction (pesticides and dioxin) or for contaminants associated with particles (atmospheric particle constituents). Finally, numerous pollutants seem to have multiple effects, and rather than being content with the first mechanism of action unveiled, we must explore the whole range of possible mechanisms (for example the metal cadmium is also an endocrine-disturbing substance). Recent studies also indicate that even when different pollutants have the same receptor, their effects may diverge in as much as the mode of activation of the receptor depends on the nature of the ligand.

In conclusion, experimental toxicology is an essential research link in the field of cancer and environment. It cannot replace epidemiology for demonstrating the relationship between pollutant and cancer pathology but can make an important contribution by establishing biological plausibility. Other applications of the experimental approach are the reinforcement of predictive toxicology, the development of exposure markers, and improved definition of sources. However, such applications must not eclipse the importance of this discipline for extending our knowledge.

Numerous questions remain to be answered through this experimental approach, particularly with respect to low doses, mixtures and complex modes of action. New technologies should help answer pending questions.

References

1. Britton M . The epidemiology of mesothelioma Semin Oncol 2002 29 18 25 . [PubMed: 11836665]
2. Bruning T , Bolt HM . Renal toxicity and carcinogenicity of trichloroethylene: key results, mechanisms, and controversies Crit Rev Toxicol 2000 30 253 285 . [PubMed: 10852497]
3. Czene K , Lichtenstein P , Hemminki K . Environmental and heritable causes of cancer among 9.6 million individuals in the Swedish family-cancer database Int J Cancer 2002 99 260 266 . [PubMed: 11979442]
4. Foulkes WD . Of mice and women Cancer Cell 2002 1 11 12 . [PubMed: 12086881]
5. Green DR , Evan GI . A matter of life and death Cancer Cell 2002 1 19 30 . [PubMed: 12086884]
6. Hecht SS . Tobacco smoke carcinogens and lung cancer J Natl Cancer Inst 1999 91 1194 1210 . [PubMed: 10413421]
7. Hemminki K , Dong C , Vaittinen P . Cancer risks to spouses and offspring in the family-cancer database Genet Epidemiol 2001 20 247 257 . [PubMed: 11180450]
8. Ince TA , Weinberg RA . Functional genomics and the breast cancer problem Cancer Cell 200 1 15 17 . [PubMed: 12086883]
9. Lichtenstein P , Holm NV , Verkasalo PK , Iliadou A , Kaprio J . et al. Environmental and heritable factors in the causation of cancer--analyses of cohorts of twins from Sweden, Denmark, and Finland N Engl J Med 200034378–85.(Commentaries N Engl J Med 2000, 343 : 1494) [PubMed: 10891514]
10. Meek ME , Bucher JR , Cohen SM , Dellarco V , Hill RN . et al. A framework for human relevance analysis of information on carcinogenic mode of action Crit Rev Toxicol 2003 33 591 653 . [PubMed: 14727733]
11. Orsulic S , Li Y , Soslow RA , Vitale-Cross LA , Silvio Gutkind J , Varmus HE . Induction of ovarian cancer by defined multiple genetic changes in a mouse model system Cancer Cell 2002 1 53 62 . [PMC free article: PMC2267863] [PubMed: 12086888]
12. Pott P . 1775 reproduced in Natl Cancer Inst Monograph 1963 10 7 .
13. Sawyers CL . Disabling Abl-perspectives on Abl kinase regulation and cancer therapeutics Cancer Cell 2002 1 13 15 . [PubMed: 12086882]
14. Tucker MA , Goldstein AM . Melanoma etiology: where are we? Oncogene 2003 22 3042 305 . [PubMed: 12789279]
15. Weber BL . Cancer genomics Cancer Cell 2002 1 37 47 . [PubMed: 12086886]

SNPs-cancer associations

Particular attention was initially given to SNPs of genes involved in the metabolism of toxic chemicals (conversion of toxic chemicals into intermediate metabolites by phase I enzymes such as cytochromes P450 and detoxification by phase II enzymes such as glutathione transferases and N-acetyltransferases). Other genes that can influence the risk of cancer, such as those involved in DNA repair, immunity, cell cycle control, or toxic substance dependency, are currently being studied. A preliminary list of candidate genes is given in Table 6.I.

Table 6.I

Preliminary list of genes that might influence the development of cancer (from Brennan, 2002).

Results of hundreds of epidemiological studies (essentially case-control studies) on various cancers have been published over the last 10 years. In spite of this sizeable effort, the results in terms of acquired knowledge are rather disappointing in that the positive associations found have not always been confirmed. The relatively small size of the populations studied (generally 100 to 300 cases) and lack of statistical power to demonstrate modest effects (ORs below 1.5) could in part explain the discordance between results. Indeed, estimates of the number of subjects necessary to detect ORs between 1.2 and 1.5 with a statistical power of 80% indicate that 500 to 2,000 cases (and an equal number of control subjects) are necessary with a prevalence of exposure to the factor under scrutiny (e.g. genetic polymorphism) of 50% in the general population (Figure 6.1) (Brennan, 2002).

Figure 6.1

Numerous synthetic studies (systematic reviews of the literature and meta-analyses of published studies) have been carried out to evaluate the relationship between specific SNPs and certain cancers (Houlston and Peto, 2004). However, few of them have made it possible to demonstrate the existence of significant associations (breast cancer and CYP2C19-(TTTA)₁₂, GSTP1-Ile105Val, and TP53-Arg72Pro; colon cancer and APC-I130K, MTHFR-Ala677Val, and HRAS-VNTR; cancer of the bladder and NAT2-slow acetylator and GSTM1 deletion; cancer of the lung and GSTM1 deletion) (Table 6.II).

Table 6.II

Significant SNPs-cancer associations demonstrated by synthetic studies (from Houlston and Peto, 2004).

Gene-environment interactions

Taking into account genetic factors in the study of cancer-environment relationships could help identify different risk levels in exposed sub-groups of individuals. A well-documented example of this is the risk of bladder cancer and exposure to aromatic amines as a function of NAT2 genotype (fast and slow acetylators). Differences in cancer risk have been observed according to NAT2 genotype; thus, not taking into account this genetic factor in studies on the relationship between cancer of the bladder and aromatic amines leads to average and therefore diluted risk estimates. A recent review of the literature (Kelada et al., 2004) indicates that a certain number of genetic polymorphisms could modify the risk of cancer associated with environmental or occupational exposure (Table 6.III).

Table 6.III

SNPs that could modify the risk of cancer associated with environmental or occupational exposure (from Kelada et al., 2004).

It is, however, necessary to point out that interaction studies between two genetic factors or between a genetic factor and an environmental factor generally require large numbers of subjects, on the order of several thousand cases and controls.

Means of calculating the interactions between two dichotomic factors (present/absent) are given in Table 6.IV, according to whether the studies involve cases with or without controls (Botto and Khoury, 2004).

Table 6.IV

Means of calculating the interactions between two dichotomic factors (from Botto and Khoury, 2004).

In conclusion, the studies carried out to date have generally analyzed a relatively restricted number of polymorphisms. Recent advances in the identification of new variants and high-throughput genotyping techniques should facilitate the simultaneous analysis of a large number of polymorphisms, single genes and multiple genes, within a single pathway. However, the simultaneous study of a large number of polymorphisms requires samples of considerable size. Furthermore, the large quantity of genotypic data generated in this way requires the development of new statistical methodology.

References

1. Benhamou S , Lee WJ , Alexandrie AK , Boffetta P , Bouchardy C . et al. Meta- and pooled analyses of the effects of glutathione S-transferase M1 polymorphisms and smoking on lung cancer risk Carcinogenesis 2002 23 1343 1350 . [PubMed: 12151353]
2. Botto LD, Khoury MJ. Facing the challenge of complex genotypes and gene-environment interaction: the basic epidemiologic units in case-control and case-only designs. In: Human genome epidemiology. Khoury MJ, Little J, Burke W, eds. Oxford University Press, 2004: 111–126.
3. Brennan P . Gene-environment interaction and aetiology of cancer: what does it mean and how can we measure it? Carcinogenesis 2002 23 381 387 . [PubMed: 11895852]
4. Dunning AM , Healey CS , Pharoah PD , Teare MD , Ponder BA , Easton DF . A systematic review of genetic polymorphisms and breast cancer risk Cancer Epidemiol Biomarkers Prev 1999 8 843 854 . [PubMed: 10548311]
5. Engel LS , Taioli E , Pfeiffer R , Garcia-Closas M , Marcus P . et al. Pooled analysis and meta-analysis of glutathione S-transferase M1 and bladder cancer: a HuGE review Am J Epidemiol 2002 156 95 109 . [PubMed: 12117698]
6. Houlston RS , Peto J . The search for low-penetrance cancer susceptibility alleles Oncogene 2004 23 6471 6476 . [PubMed: 15322517]
7. Houlston RS , Tomlinson IP . Polymorphisms and colorectal tumor risk Gastroenterology 2001 121 282 301 . [PubMed: 11487538]
8. Kelada SN, Eaton DL, Wang SS, Rothman NR, Khoury MJ. Applications of human genome epidemiology to environmental health. In: Human genome epidemiology. Khoury MJ, Little J, Burke W, eds. Oxford University Press, 2004: 145–167.
9. Kinzler KW , Vogelstein B . Cancer-susceptibility genes. Gatekeepers and caretakers Nature 1997 386 762 763 . [PubMed: 9126728]
10. Marcus PM , Vineis P , Rothman N . NAT2 slow acetylation and bladder cancer risk: a meta-analysis of 22 case-control studies conducted in the general population Pharmacogenetics 2000 10 115 122 . [PubMed: 10761999]
11. Nagy R , Sweet K , Eng C . Highly penetrant hereditary cancer syndromes Oncogene 2004 23 6445 6470 . [PubMed: 15322516]

Dose-response relationship, an element that determines priorities

Concluding that there is a causal link between exposure to a given agent and the occurrence of a disease is a necessary but not sufficient condition to determine action priorities. To do this, one also needs to know who is exposed and to what extent, which involves descriptive epidemiological studies, and to determine the effect of a given amount of exposure or dose of a particular carcinogen, i.e., the dose-response relationship.

Mutagenic or genotoxic carcinogens

The most distinguished international committee experts, particularly those dealing with radioprotection issues (International Commission on Radiological Protection) consider that the mode of action of genotoxic carcinogens (mutagenic or responsible for chromosomal abnormalities) does not involve a threshold. The arguments in favor of this hypothesis are presented below.

According to this model, any dose of a genotoxic carcinogen is responsible for an excess of cancer risk. One of the major challenges then is to determine the shape of the dose-response curve at low doses. Indeed, the majority of the population is usually exposed to low doses. The problem is that the effects of low doses are very subtle and cannot be observed. One must therefore extrapolate from observable data what is likely to occur in the non-observable domain (Figure 7.1). To do this, one can use epidemiological data when it is available, which is rather rare, or experimental data on animals, which is more frequently the case. Either way, different mathematical models of extrapolation can generally adjust satisfactorily to the data in the observable domain (Figure 7.2). The quality of such adjustment is judged on the basis of a statistical test of adequateness. However, the shape of the dose-response curve outside the observable domain is both determined on the basis of the observable data used by the models and impossible to validate. Figure 7.3 shows the results of animal data extrapolation for the risk of death by cancer after dioxin (2,3,7,8-tetrachlorodibenzo-p-dioxin) administration. Although several models adequately adjust to the data, the risk estimates provided by each of them for a given dose vary widely, particularly as the dose level is reduced. In the first instance, the United States Environmental Protection Agency (U.S. EPA) proposes using the so-called linearized multistage or LMS model, provided it is sufficiently adequate for the data used, for two reasons. First, because it is partially based on biological considerations in accordance with current ideas on carcinogenesis mechanisms; and second, because it generally gives rise to the most pessimistic risk estimates, which affords extra precaution. It is worth noting that this is not the case for the example given in Figure 7.3 for curves labeled MS or MS₂, the latter case corresponding to a truncated set of data in order to ensure better adequateness.

Figure 7.1

Figure 7.2

Diagram of the possible extrapolations from measured risks at very low doses (from Brenner et al., 2003). Central tendency risk measurements are represented here by black circles with the corresponding confidence intervals. All the models used fit the (more...)

Figure 7.3

Applying different extrapolation models to a set of animal data (from Longstreth and Hushon, 1983). Excess death risk by cancer in the case of 2,3,7,8-tetrachlorodibenzo-p-dioxin M: Multi-hit; W: Weibull; L: Logit; MS: Linear multistep; P: Probit; MS (more...)

In conclusion, the choice of model is a parascientific decision that involves considerations such as principles (to be as protective as possible), knowledge or hypotheses on mechanisms of action, and mathematical modelization tools.

Non-mutagenic or non-genotoxic carcinogens

For carcinogens intervening at stages other than the initial stage of mutagenesis (e.g. at the promotion or progression stage), it is now accepted that there is a dose threshold below which there is no effect. The value of the threshold now remains to be determined. This can sometimes be done on the basis of epidemiological studies (Figure 7.4). However, it is more often than not obtained from experimental studies on animals such as rodents. For a given effect, the value of the threshold dose of the agent responsible varies according to experimental conditions. On the one hand, the species tested and the particular strain used have a specific sensitivity to the tested carcinogen. On the other hand, the statistical power of the study influences the dose threshold observed, which can be lower if the statistical power is increased by using more animals per dose group. It is therefore necessary to judge in this case which value ought to be chosen and whether the transposition of animal data to the human species is plausible. Once again, one is faced with having to pass judgment and make a decision on the basis of scientifically established data but under circumstances involving a measure of uncertainty.

Figure 7.4

Risk projections

Observations can also permit impact estimates by testing one or more models that will serve to predict the shape of the dose-response curve as a function of the dose received as well as the incidence of and mortality by a given cancer in the population in the absence of exposure (baseline risk). This approach can be useful to estimate lifelong risk in a cohort for which one wishes to determine risk after exposure while significant numbers of the subjects followed are still alive and free of cancer. The projection model may be multiplicative (relative risk model) with respect to the baseline risk or additive (constant risk model). The exposure impact is clearly different according to the model chosen (Figure 7.5).

Figure 7.5

Overall effect of the existence of sensitive sub-populations on the dose-response relationship

For breast cancer linked to ionizing radiation exposure (Brenner et al., 2003), there is among the general population a very small proportion (about 0.25%) of women that are highly sensitive, for genetic reasons in particular, to the effect of genotoxic carcinogens, yet the dose-response relationship is linear at low doses for these subjects just as it is for the general population. When the dose increases, the number of cancers among sensitive women reaches a plateau (all of these women will develop cancer). For the two populations combined, the curve will be overall supralinear (Figure 7.6). In this case, applying a linear relationship for the exposed population as a whole, including highly sensitive women, will underestimate the risk.

Figure 7.6

Transpositions, extrapolations, analogies

In the case where risk is projected, if the risk projection extrapolated from last follow-up data is compatible with a proportional risk model, the estimated risk for a given dose is going to depend on baseline incidence or mortality, which varies from one population to another and has a noticeable influence. Table 7.II helps us appreciate the problem by giving a few examples of differences in death rates between countries for three cancer localizations.

Table 7.II

Standardized death rates per 100.000 for various cancers in 1988.

Indeed, the general issue is whether one can justifiably transpose observations made on one population to another population in the observable domain. For example, is the dose-response relationship observed for breast cancer risk due to ionizing radiation in survivors of the Hiroshima and Nagasaki atomic bombing valid to estimate the exposure risk of the French population to the same radiation 50 years later? There is no simple answer to this question except to say that the further away we move in time or space, the greater the uncertainty. Once more, it is necessary to exercise judgment because elements confirming the comparability of different populations are lacking.

Lack of epidemiological data makes the problem even more acute when transpositions must be made from animals to humans. Knowledge about mechanisms of action and their phylogenetic conservation is, in this case, an important, although rarely decisive, guide.

Finally, a particular variety of transposition is what Hubert (2003) in the wake of Hill (1965) calls an analogy, when one supposes, for example, that an effect observed after exposure by inhalation might also occur when the same agent is ingested. Figure 7.7 summarizes the various transpositions discussed.

Figure 7.7

Example of possible gain from transposition/extrapolation/analogy in the case of radon exposure (from Hubert, 2003). Inferences from data on French uranium miners for the French population as a whole

Controversy about linearity without a threshold

The existence of DNA repair mechanisms is well established. To put it simply, a mutation can only be stable when there is concomitant damage to the two complementary DNA strands. At very low dose, some authors argue the probability of such concomitant damage is nil because the agent concentration at the relevant chromosome site is bound to be insufficient, which means there must always be a threshold. There is a controversy as to whether such conclusions are justified (Brenner et al., 2003), but the pros and cons cannot be discussed in detail here. In any case, the problems of extrapolation and transposition discussed above are also relevant here. In other words, demonstrating the existence of a cellular or pluricellular threshold poses the problem of knowing if such a threshold exists at the organism and population level (in the case of highly sensitive populations) and if so, of determining its value. We are far from being able to estimate such a threshold on solid grounds, and there is really no operational alternative to the pragmatic approach, which uses the linearity model without a threshold.

In conclusion, the shape of the dose-response curve contributes to making a causality judgment and, if the latter is deemed justified, determining sanitary impact and action priorities, provided the exposure distribution is known for the target population. In any case, this approach deals with single carcinogens and cannot in practice take into account multiple exposures. It is also important to emphasize that qualitative and quantitative uncertainties are such that evaluation of the dose-response relationship, although based as much as possible on well-established scientific facts, remains a question of judgment.

References

1. Bertin M. Les effets biologiques des rayonnements ionisants. Electricité de France/SODEL, Paris, 1991.
2. Brenner DJ , DollL R , Goodhead DT , Hall EJ , Land CE . et al. Cancer risks attributable to low doses of ionizing radiation: assessing what we really know Proc Natl Acad Sci USA 2003 100 13761 13766 . [PMC free article: PMC283495] [PubMed: 14610281]
3. Hill AB . The environment and disease: association or causation? Proc R Soc Med 1965 58 295 300 . [PMC free article: PMC1898525] [PubMed: 14283879]
4. Holder J , Menzel H . Analysis of 2,3,7,8-TCDD tumor promotion activity and its relationship to cancer Chemosphere 1989 19 861 868 .
5. Hubert P . Pour un meilleur usage du risque attribuable en santé environnementale Environnement, Risques & Santé 2003 2 266 278 .
6. Kociba RJ , Keyes DG , Beyer JE , Carreon RM , Wade CE . et al. Results of a two-year chronic toxicity and oncogenicity study of 2,3,7,8-tetrachlorodibenzo-p-dioxin in rats Toxicol Appl Pharmacol 1978 46 279 303 . [PubMed: 734660]
7. Longstreth JD, Hushon JM. Risk assessment for 2,3,7,8-tetrachlorodibenzo-p-dioxin(TCDD), In: Human and environmental risks of chlorinated dioxins and related compounds. Tucker RE, Young AL, Gray AP, eds. Plenum Press, New York, 1983: 639–666.

Definition and generalities

Attributable risk is defined by the following ratio:

The numerator is equal to the difference between two probabilities, the overall probability of the disease, Pr(D), in the population (generally made up of both exposed, E, and unexposed subjects, Ē) and the hypothetical probability of the disease in the same population supposing exposure was totally eliminated, Pr(D | Ē). It therefore quantifies the additional probability of the disease in the population associated with the presence of the exposure factor. The attributable risk AR measures the corresponding proportion relative to the overall probability of disease in the population. Such different probabilities correspond to disease risks (probabilities of developing the disease in the course of a given lapse of time) but, under certain conditions or in the case of certain applications, disease incidence rates may be substituted to risks to define the attributable risk.

Others formulations as a function of the prevalence of exposure and the strength of association

Unlike measures of association such as relative risk (ratio of disease risks in exposed and unexposed subjects), attributable risk depends both on the strength of association between the exposure factor and the disease and the prevalence of this factor in the population, p _E . This joint dependency becomes apparent if the formula (1) is rewritten as follows. Pr(D) can be expressed as:

in both the numerator and denominator. Now,

where RR represents the relative risk (or the ratio of incidence rates of the disease in exposed and unexposed subjects). Because the term Pr(D|Ē) cancels out, the attributable risk AR can be written as follows (Cole and MacMahon, 1971; Miettinen, 1974):

i.e., both as a function of the exposure prevalence in the population, p _E , and of relative risk (or ratio of incidence rates), RR.

An alternative formulation emphasizes this joint dependency. By expressing Pr(D) and Pr(D|E) as shown above, the numerator of formula (1) can be rewritten as follows:

Using Bayes's theorem, one can express Pr(D|E) as Pr(E|D) Pr(D) /p _E , which leads to the following expression for the numerator of formula (1):

thus giving the following (Miettinen, 1974):

i.e., both a function of the exposure prevalence in subjects suffering from the disease under consideration (cases), p _E|D , and of relative risk (or ratio of incidence rates), RR.

Thus, according to the exposure prevalence, a high relative risk value may give rise to a high or low attributable risk value, which leads to very different consequences in terms of public health. One of the consequences is that attributable risk is generally not portable from one population to another because exposure prevalence can vary extensively between different populations in time and space. This situation differs from that observed for measures of association, which are generally much more readily portable from one population to the next because the strength of an association is in most cases subject to little variation between populations except when the exposure factor strongly interacts with environmental or genetic factors specific to the various populations.

Use and interpretation

Attributable risk is used to quantify the impact of primary prevention policies relative to the disease under study or to quantify the fraction of disease that can be explained by known factors. There follows two types of interpretations.

Properties

Two important properties of attributable risk deserve to be described: dependence on the definition of the reference level, and distributivity.

Dependence on the definition of the reference level

The values of attributable risk strongly depend on the definition of the reference level for exposure. This level corresponds either to zero exposure or to an appropriate baseline level in case of an exposure continuum. A more restrictive definition of the reference level leads to a higher proportion of individuals being considered as exposed. Thus, the narrower the reference level by dropping the more exposed subjects (and therefore theoretically more at risk) from it, the higher the value of attributable risk. This property has a major impact on estimates of attributable risk as Benichou (1991) and Wacholder et al. (1994) have shown. For example, Benichou (1991) obtained an estimate of the risk of esophagus cancer attributable to alcohol consumption higher than or equal to 80 g/day (i.e. reference level of 0 to 79 g/day) of 38% for the Île-et-Vilaine department in France. This estimate went up to 70% for alcohol consumption higher than or equal to 40 g/day (using a reference level of 0 to 39 g/day). This property plays a role when studying a continuous exposure with a continuous risk gradient and no obvious threshold. Thus, estimates of attributable risk can be interpreted validly and make sense only by reference to a clearly defined baseline level.

The above estimates of attributable risk for esophagus cancer in Île-et-Vilaine were obtained from data of a case-control study carried out in the 1970s. This study included 200 cases of esophagus cancer and 775 controls selected on a random basis on the department's electoral listings (Tuyns et al., 1977). Assessment of the associations of alcohol consumption and smoking with esophagus cancer is illustrated in detail in a reference publication by Breslow and Day (1980) that presents several approaches to estimate the odds ratio with or without age adjustment. In the present illustration and in the course of previous work (Benichou, 1991), four levels of alcohol consumption (0–39, 40–79, 80–119, and ≥120 g/day), three levels of tobacco consumption (0–9, 10–29, ≥30 g/day) and three age groups (25–44, 45–54, ≥55) are being considered. There were 29, 75, 51, and 45 cases with a respective alcohol consumption of 0–39, 40–79, 80–119, and ≥120 g/day. The numbers of corresponding control subjects were 386, 280, 87, and 22, respectively.

The first reference level considered (0–79 g/day) includes 104 cases and 666 controls, which leaves 96 cases and 109 controls in the exposed category (≥80 g/day) (Table 8.I). The crude (i.e. non adjusted) odds ratio is equal to 5.6 (96 × 666/104 × 109). According to the methods described below, the crude attributable risk is estimated at 39.5% for alcohol consumption and about 38% after adjusting for age and tobacco consumption. The second reference level considered (0–39 g/day) is more restrictive and only includes 29 cases and 286 controls, which leaves 171 cases and 489 controls in the exposed category (≥40 g/day) (Table 8.I). The corresponding crude odds ratio is estimated at 5.9 (171 × 386/29 × 389). According to the methods described below, the crude attributable risk is 70.9% and the adjusted attributable risk lies between 70 and 72%. This substantial increase results mainly from the higher proportion of exposed subjects with this more restrictive definition of the reference category (50% instead of 14% of exposed controls).

Table 8.I

Case-control study on esophagus cancer; numbers of cases and controls in the reference category and in the category exposed to daily alcohol consumption for two definitions of the reference category (from Tuyns et al., 1977).

Case of multiple exposure factors

Attributable risk is frequently estimated in multifactorial situations when trying to evaluate the joint and individual impact of multiple exposure factors associated with disease occurrence. This raises a problem because individual contributions of the factors to attributable risk are non-additive.

Non-additivity

When considering multiple exposure factors, it is possible to estimate distinct attributable risks for each exposure factor as well as an overall joint attributable risk for the various exposure factors. Walter (1983) has shown that the sum of risks attributable to each factor is not equal to the joint attributable risk unless two specific conditions are fulfilled: there is no joint exposure to the different exposure factors in the population (e.g., to tobacco and alcohol), and the effects of the exposure factors on disease risk are additive. For two exposure factors, the latter condition means that the relative risk for exposure to the two factors, RR₁₂, is linked to the relative risks for exposure to factors 1 and 2 separately, RR₁ and RR₂, respectively, by the formula (RR₁₂-1) = (RR₁ - 1) + (RR₂ - 1). If none of the two conditions above are verified, then the sum of the risks attributable to each factor differs from the joint attributable risk, and the difference can be substantial.

Table 8.II, taken from Begg (2001), illustrates this problem. The table considers two exposure factors E ₁ and E ₂ with a prevalence in the population of 0.25 for each of the joint exposure categories. Each of these exposure factors multiplies disease risk by 9 with a joint multiplicative effect such that the risk is multiplied by 81 in the case of joint exposure to both factors. By using either formula (1) or formula (2), one obtains an attributable risk of 80% for both factors E ₁ and E ₂ separately. Indeed, with formula (2) for example, the attributable risk for factor E ₁ is:

i.e., AR₁ = 80%. The same applies to factor E ₂ because the problem is perfectly symmetrical in this particular case. Thus, the sum of the separate attributable risks for factors E ₁ and E ₂, i.e., AR₁ + AR₂, cannot be equal to the joint attributable risk for factors E ₁ and E ₂ because the sum is greater than 100%! The joint attributable risk for factors E ₁ and E ₂ can be obtained by using formula (1):

which is equivalent to:

where Pr(D | Ē) is the risk of developing the disease in subjects that are neither exposed to E ₁ nor E ₂, i.e., 0.01. The probability Pr(D) represents the risk of developing the disease in the population so that when the joint prevalence of exposure factors E ₁ and E ₂ are taken into account, the probability is:

and, the joint attributable risk for factors E ₁ and E ₂ is:

i.e. AR₁₂ = 96%, which is clearly lower than the sum AR₁ + AR₂. The non-additivity problem comes from the fact that by forming the sum AR₁ + AR₂, one is not considering the same reference levels as when considering the joint attributable risk AR₁₂. For the latter, the reference level is the category that corresponds to an absence of exposure to E ₁ and E ₂. In the case of the risk attributable to factor E ₁, i.e., AR₁, the reference level corresponds to an absence of exposure to E ₁ only and therefore includes subjects both exposed and unexposed to E ₂ (in equal proportion in this example). Similarly, for the risk attributable to factor E ₂, i.e., AR₂, the reference level corresponds to an absence of exposure to E ₂ only and therefore includes subjects both exposed and unexposed to E ₁ (in equal proportion in this example). This means that the contribution of the category of subjects exposed to both E ₁ and E ₂ is taken into account more than once in the sum AR₁ + AR₂, which explains the inadequateness of calculating AR₁ + AR₂ except in the specific cases described by Walter (1983).

Table 8.II

Illustration of the phenomenon of non-additivity of attributable risks for two exposure factors E1 and E2 and multiplicative risks.

Solutions proposed – Sequential attributable risk

Because the non-additivity property is somewhat contrary to intuition and might engender erroneous interpretations, three alternative approaches have been suggested. The first approach is based on methods of variance decomposition (Begg et al., 1998) rather than estimating attributable risk and is therefore not directly applicable here. The second approach is based on estimating the probability of causation of each exposure, which is relevant in legal or compensation procedures where the aim is to determine the probability that the disease of a given individual having been exposed to several risk factors is attributable to one of these factors in particular, for example the probability that the lung cancer of a smoker occupationally exposed to asbestos is attributable to occupational exposure to asbestos (Cox 1984, 1985; Lagakos and Mosteller, 1986; Seiler, 1986; Seiler and Scott, 1987; Benichou, 1993; McElduff et al., 2002; Llorca and Delgado-Rodriguez, 2004). This approach being individual and non populational is not directly useable here.

The third approach is based on an extension of the concept of attributable risk (Eide and Gefeller 1995; Land et al., 2001) and is therefore more immediately relevant to the situation under consideration here. It relies on partition techniques (Land and Gefeller, 1997; Gefeller et al., 1998) and conserves the conceptual framework of attributable risk estimation while introducing the notion of sequential attributable risk, which generalizes the concept of attributable risk. The principle of this approach is to determine an order for the exposures under consideration. The contribution of each exposure is then evaluated sequentially according to its ranking. The contribution of the exposure that is placed first is calculated like the standard attributable risk for this exposure, were it separate. The contribution of the second exposure is obtained by calculating the difference between the joint attributable risk estimated for the first two exposures and the attributable risk estimated for the first exposure singly. Similarly the contribution of the third exposure is the difference between the joint attributable risk estimated for the first three exposures and the joint attributable risk for the first two exposures, etc. A multidimensional vector representing the contributions of each exposure is thus obtained.

Estimating these contributions is useful for potential prevention programs that consider the successive rather than simultaneous eliminations of exposures from the population. Thus, each step provides information about the additional effect of eliminating a given exposure after having eliminated the preceding exposures under consideration. After a certain number of steps, additional effects can become very small, thereby indicating that there are no reasons to consider additional steps. By construction, the sum of the contributions of each factor is equal to the joint attributable risk for all the exposures considered, which eliminates the problem of non-additivity. The contribution of each exposure factor depends on the order of the exposures considered. The most useful orders to consider depend on practical possibilities of implementation of potential prevention programs in a given population. Mean contributions, called partial attributable risks, can be obtained for each exposure factor by calculating the mean contribution for all the orders possible (Eide and Gefeller, 1995). Methods for visualizing sequential and partial attributable risks have been developed by Eide and Heuch (2001). These are illustrated in Figure 8.1. Land et al. (2001) present a detailed review of the properties and interpretation of sequential and partial attributable risks.

Figure 8.1

Taking once again cancer of the esophagus as an example, tobacco consumption is a well-known major risk factor for this cancer in addition to alcohol. It is therefore important to estimate the joint impact of tobacco and alcohol consumption on the occurrence of cancer of the esophagus relative to the impact of alcohol consumption alone. By using the first category (0–9 g/day) as the reference level category of tobacco consumption, we have 78 cases in the tobacco consumption reference level, 122 cases in the exposed category (i.e. ≥10 g/day), 447 controls in the reference category and 328 controls in the exposed category. From this data, the crude odds ratio for tobacco consumption of at least 10 g/day is estimated at 2.1 and the crude attributable risk at 32.4%. Because there are 9 cases and 252 controls in the category jointly exposed to tobacco and alcohol (0–39 g/day of alcohol and 0–9 g/day of tobacco), the crude joint odds ratio is estimated at 10.2 and the crude joint attributable risk for alcohol consumption of at least 40 g/day or tobacco consumption of at least 10 g/day is estimated at 86.2%.

Furthermore, the crude attributable risk for alcohol consumption of at least 40 g/day is estimated at 70.9% (see above). If one considers first the elimination of alcohol consumption of over 39 g/day, followed by the elimination of tobacco consumption of over 9 g/day, the sequential attributable risk is estimated at 70.9% for a high daily consumption of alcohol and 86.2 − 70.9 = 15.3% for substantial tobacco consumption, once daily alcohol consumption has been eliminated. The additional impact of eliminating substantial tobacco consumption therefore appears to be rather limited (Figure 8.1 a). If we consider the other order possible, i.e. elimination of substantial tobacco consumption first, the sequential attributable risk is estimated at 32.4% for substantial tobacco consumption and 86.2 − 32.4 = 53.8% for high alcohol consumption once substantial tobacco consumption has been eliminated. Thus the major additional impact of eliminating the high daily consumption of alcohol remains (Figure 8.1b). These results are summarized by partial attributable risks for high alcohol consumption and substantial tobacco consumption of 62.4% and 23.9%, respectively, which reflects the greater impact of alcohol consumption on the occurrence of esophagus cancer.

Estimating attributable risk

Attributable risk can be estimated from the main types of epidemiological studies.

Basic principles for estimating attributable risk

Attributable risk can be estimated from cohort studies because all the terms in formulae (1), (2), and (3) can be obtained directly from such studies. In case-control studies, risks or incidence rates specific for each exposure level are not available unless the data are complemented by follow-up or incidence data. One must therefore rely on odds ratio estimates, using formula (2), and estimate the prevalence of exposure in the population, p _E , from the proportion of exposed subjects among controls, by hypothesizing a rare disease (a hypothesis also used in odds ratio rather than relative risk estimation). To estimate crude attributable risks, the odds ratio is calculated with the formula ad/bc and p _E by c/(c + d), where a, b, c, and d are the numbers of exposed cases, unexposed cases, exposed controls, and unexposed controls, respectively. Alternatively, one can use formula (3), in which prevalence of exposure in subjects affected by the disease, p _{E |D}, can be directly estimated from the cases by the ratio a/(a+b) and the relative risk by the odds ratio (ad/bc) as before. Whichever equation is used, the crude attributable risk estimate is finally obtained by the formula (ad − bc)/{d(a + b)}.

Estimators of variance are available that allow one to obtain confidence intervals for attributable risk. The respective merits of confidence intervals, whether or not based on mathematical transformations (logarithmic or logistic) have been discussed in the literature (Walter, 1975, Leung and Kupper, 1981, Whittemore, 1982). Detailed reviews of the basic principles involved in the calculation of attributable risk for different types of epidemiological studies have been presented (Walter, 1976, Benichou, 2000a, 2001).

Taking once again the example of esophagus cancer with the most restrictive definition of the reference category of daily alcohol consumption (0–79 g/day), the crude attributable risk is calculated as follows:

i.e., 70.9% with a standard deviation of 0.051 or 5.1%. The corresponding 95% confidence intervals for attributable risk are [60.9–80.9%], [58.9–79.4%], and [60.0–79.8%] with no transformation, with the logarithmic transformation and logistic transformation, respectively, thus giving three very similar results in this particular example.

Adjusted estimation – Stratification and regression methods

Several approaches have been proposed for adjusted AR estimation. The most general approach is based on the use of regression models. It relies on expressing attributable risk with the following formula (Bruzzi et al., 1985; Benichou, 1991):

Each ρ _ij term represents the proportion of individuals with the disease (cases) with a level of exposure i (i = 0 for the reference level, i = 1,…, I for exposed levels) and a level of adjustment factor j (confounding factors). These terms can be calculated from cohort or case-control studies data by using the observed proportion of exposed subjects among cases. Each RR _{i |j} ⁻¹ term represents the inverse of the relative risk, of the incidence rate ratio or of the odds ratio according to the context, for a level of exposure i at a given level j of adjustment factors. These terms can be calculated using regression models from cohort or case-control data. According to the type of study, Poisson or conditional or unconditional logistic regression models can be used. Thanks to these models, attributable risk calculations are adjusted for confounding factors and can also include terms representing the interaction between such confounding factors and the exposure factor(s) under consideration. Specific variance estimators are available that allow calculation of confidence intervals for attributable risks (Benichou and Gail, 1989, 1990).

Such a regressional approach not only includes the non-adjusted approach as a specific case and the two adjusted approaches based on stratification methods but also offers additional options (Benichou, 1991). The non-adjusted approach corresponds to RR _{i |j} ⁻¹ models that do not take confounding factors into account. Mantel-Haenszel's stratified approach (Greenland, 1984, 1987; Kuritz and Landis, 1987, 1988a, b) corresponds to models that include exposure and confounding factors but no terms to account for interaction between exposure and confounding factors. The weighted sum stratified approach (Walter, 1976; Whittemore, 1982, 1983) corresponds to saturated models where all the terms to account for possible interactions between exposure and confounding factors are included. Additionally, intermediate models can be considered that take into account the interaction between exposure and a single confounding factor (rather than all of them), or models in which the main effects of several confounding factors are not modeled by a saturated approach.

Coming back to the example of esophagus cancer with the most restrictive definition of the reference category for daily consumption of alcohol, an unconditional logistic model with two parameters, a general "intercept" parameter and a parameter for high alcohol consumption, can be used that does not take into account tobacco consumption or age. The calculated non-adjusted odds ratio is 5.9 as above. The formula given above for 1-AR is reduced in this case to a single two-term sum (i = 0.1) corresponding to the unexposed and exposed categories. The non-adjusted attributable risk is estimated at 70.9% (with a standard variation of 5.1%), identical to the crude attributable risk calculated above. By adding eight terms for tobacco consumption and age in the logistic model, the adequateness of the model is significantly improved (p<0.001) and allows the calculation of an adjusted odds ratio of 6.3 and an adjusted attributable risk of 71.6% (with a standard variation of 5.0%), similar to what would be obtained with Mantel-Haenszel's stratification approach. By adding two terms for the interaction of tobacco and alcohol consumption, the attributable risk calculated is slightly lower (70.3% with a standard variation of 5.4%). Finally, by adding six more parameters that take into account one-to-one interactions of alcohol consumption with the joint level of tobacco consumption and age, one obtains a completely saturated model in which nine different odds ratios for alcohol consumption are calculated as in the weighted sum stratified approach. Thus the attributable risk is only slightly modified (with a value of 70.0%, identical to the value obtained with the weighted sum stratified approach, and a standard variation of 5.6%).

Detailed reviews on adjusted attributable risk estimation are available in the literature (Benichou, 1991, 2001; Gefeller, 1992; Coughlin et al., 1994).

Other measures of impact

Apart from attributable risk, other measures of impact have been proposed.

Preventable and prevented fractions

When considering a protective exposure or preventive intervention, an appropriate alternative to the attributable risk is the prevented or preventable fraction, PF, defined as the ratio (Miettinen, 1974):

where Pr(D) is the probability of occurrence of the disease in a population that comprises exposed (E) and unexposed individuals (Ē), and Pr (D | Ē) is the hypothetical probability of the disease in the same population if protective exposure were to be absent. According to the situation, such probabilities either refer to the risk or to the rate of disease incidence.

The prevented or preventive fraction PF can be written:

which is a function of the prevalence of exposure, p _E , and the relative risk. Thus, as in the case of attributable risk, a strong association between exposure and disease can correspond to a high or low value of the prevented or preventive fraction PF because it is dependent on the prevalence of exposure. As a result, the PF fraction is not usually transposable from one population to another. Again as in the case of attributable risk, it can be useful to compare the estimated value for the PF fraction in different subgroups of subjects to target prevention efforts on sub-groups for which impact is likely to be the greatest.

For a protective factor (RR<1), the value of the PF fraction lies between 0 and 1 and increases with the prevalence of exposure and strength of association between exposure and disease.

The PF fraction measures the impact of an association between a protective exposure and disease at the whole population level. Interpretation in terms of public health bears on the proportion of avoided cases (prevented fraction), thanks to the introduction of a protective exposure or preventive intervention in the population, among the totality of cases that would have arisen in the absence of such protective exposure or preventive intervention (a posteriori evaluation). Furthermore, it is possible to evaluate prevention programs a priori by measuring the proportion of cases that can potentially be avoided (preventable fraction) if a protective exposure or preventive intervention were introduced de novo in the population (Gargiullo et al., 1995). However, such interpretations are subject to the same limitations as interpretations of attributable risk.

Attributable risk AR and the PF fraction are mathematically interdependent (Walter, 1976) as shown in the equation:

This equation shows that, for a protective factor, the PF fraction generally differs from the attributable risk AR calculated by reversing the exposure coding. This is in agreement with the respective definitions of the AR and PF quantities. Although the reverse coded attributable risk measures the potential reduction in the risk of disease occurrence were all the subjects in the population to become exposed, the PF fraction measures the reduction in the potential risk of disease occurrence that results from the introduction of exposure in a population that is initially not exposed (Bénichou, 2000b).

From equation (6), it becomes apparent that calculation of the PF fraction raises identical problems to those posed by attributable risk. In particular, methods of adjusted estimation based on Mantel-Haenszel's (Greenland, 1987) and weighted sum (Gargiullo et al., 1995) stratification approaches have been proposed.

Generalized impact fraction

The generalized impact fraction (GIF) or generalized attributable fraction has been defined by Walter (1980) and Morgenstern and Bursic (1982) as being the ratio:

where the terms Pr (D) and Pr*(D) correspond to the probabilities of disease in the population for current and modified distributions of exposure respectively. As with attributable risk and the prevented or preventable fraction, these probabilities can either describe the risk of the disease or its rate of incidence, according to the situation.

The generalized impact fraction not only depends on the association between exposure and disease and on the current distribution of exposure (rather than the sole prevalence of exposure) but also on the modified distribution Pr*(D) under consideration. It is therefore a general measure of impact that includes attributable risk and the PF fraction as specific cases. Attributable risk deals with the difference between the current distribution of exposure and a modified distribution defined by the absence of exposure. The PF fraction deals with the difference between a distribution defined by an absence of exposure and the current distribution of exposure in the population (prevented fraction) or an absence of current exposure and a target distribution of exposure (preventable fraction).

The generalized impact fraction can be interpreted as the fractional reduction of disease incidence that would result from a change in the current distribution to a modified distribution of exposure in the population. Thus, it is useful to evaluate prevention programs or interventions that target either all subjects or only subjects at specific levels of exposure and are aimed at modifying or reducing exposure without necessarily eliminating it totally. For example, some interventions might only focus on heavy smokers rather than all smokers. The attributable risk corresponds to the specific case of exposure elimination by considering a modified distribution of exposure reduced to a single point, with all subjects becoming unexposed. Furthermore, the generalized impact fraction can be used to evaluate the increase in disease incidence that results from exposure modifications in a population, such as, for instance, the increase in breast cancer incidence resulting from late maternity in more recent age cohorts (Kleinbaum et al., 1982). The limitations of such interpretations are the same as for attributable risk and prevented and preventable fractions.

The generalized impact fraction has been used, for example, by Lubin and Boice (1989), who have measured the impact of a modification in radon exposure distribution on lung cancer by truncating the distribution according to different thresholds, and by Wahrendorf (1987), who examined the impact of various changes in nutritional habits on colorectal and stomach cancers.

The estimation problems and methods are similar to those that apply to attributable risk and the PF fraction. However, a possible difference lies in the necessity that might arise to consider the continuous nature of some exposures rather than rely on categories of exposure to define the modification of the distribution under scrutiny, e.g., an identical exposure reduction at each level (Benichou and Gail, 1990). Drescher and Becher (1997) have proposed an extension of the adjusted estimation approach based on regression models (Bruzzi et al., 1985; Greenland and Drescher, 1993) to estimate the generalized impact fraction in case-control studies and have considered continuous as well as categorical exposures.

References

1. Ames BN , Gold LS , Willett WC . The causes and prevention of cancer Proc Natl Acad Sci USA 1995 254 1131 1138 . [PMC free article: PMC41674] [PubMed: 7777494]
2. Begg CB . The search for cancer risk factors: when can we stop looking? Am J Public Health 2001 91 360 364 . [PMC free article: PMC1446611] [PubMed: 11236398]
3. Begg CB , Satagopan JM , Berwick M . A new strategy for evaluating the impact of epidemiologic risk factors for cancer with applications to melanoma J Am Stat Assoc 1998 93 415 426 .
4. Benichou J . Methods of adjustment for estimating the attributable risk in case-control studies: a review Stat Med 1991 10 1753 1773 . [PubMed: 1845304]
5. Benichou J . Methods of adjustment for estimating the attributable risk in case-control studies: a review (Letter) Stat Med 1993 12 94 96 . [PubMed: 8318079]
6. Benichou J. Attributable risk. In: Encyclopedia of epidemiologic methods. Gail MH, Benichou J., eds, Wiley, Chichester, 2000a: 50–63.
7. Benichou J. Preventable fraction. In: Encyclopedia of epidemiologic methods. Gail MH, Benichou J., eds, Wiley, Chichester, 2000b: 736–737.
8. Benichou J . A review of adjusted estimators of the attributable risk Stat Methods Med Res 2001 10 195 216 . [PubMed: 11446148]
9. Benichou J , Gail MH . Variance calculations and confidence intervals for estimates of the attributable risk based on logistic models Biometrics 1990 46 991 1003 . [PubMed: 2085643]
10. Breslow NE, Day NE. Statistical methods in cancer research. Vol 1: The analysis of case-control studies. International Agency for Research on Cancer Scientific Publications 1980, No. 32, Lyon. [PubMed: 7216345]
11. Bruzzi P , Green SB , Byar DP , Brinton LA , Schairer C . Estimating the population attributable risk for multiple risk factors using case-control data Am J Epidemiol 1985 122 904 914 . [PubMed: 4050778]
12. Colditz G , Dejong W , Hunter D , Trichopoulos D , Willett W . Harvard report on cancer prevention, Vol. 1 Cancer Causes Control 1996 7 S3 S59 .
13. Colditz G , Dejong W , Hunter D , Trichopoulos D , Willett W . Harvard report on cancer prevention, Vol. 2 Cancer Causes Control 1997 8 S1 S50 .
14. Cole P , MacMahon B . Attributable risk percent in case-control studies Br J Prev Soc Med 1971 25 242 244 . [PMC free article: PMC478665] [PubMed: 5160433]
15. Coughlin SS , Benichou J , Weed DL . Attributable risk estimation in case-control studies Epidemiol Rev 1994 16 51 64 . [PubMed: 7925727]
16. COX LA . Probability of causation and the attributable proportion of risk Risk Anal 1984 4 221 230 .
17. COX LA . A new measure of attributable risk for public health applications Manage Sci 1985 7 800 813 .
18. Doll R, Peto R. The causes of cancer. Oxford University Press, New York, 1981.
19. Drescher K , Becher H . Estimating the generalized attributable fraction from case-control data Biometrics 1997 53 1170 1176 . [PubMed: 9290235]
20. Eide GE , Gefeller O . Sequential and average attributable fractions as aids in the selection of preventive strategies J Clin Epidemiol 1995 48 645 655 . [PubMed: 7730921]
21. Eide GE , Heuch I . Attributable fractions: fundamental concepts and their vizualization Stat Methods Med Res 2001 10 159 193 . [PubMed: 11446147]
22. Gargiullo PM , Rothenberg R , Wilson HG . Confidence intervals, hypothesis tests, and sample sizes for the prevented fraction in cross-sectional studies Stat Med 1995 14 51 72 . [PubMed: 7701158]
23. Gefeller O . Theory and application of attributable risk estimation in cross-sectional studies Statistica Applicata 1990 2 323 331 .
24. Gefeller O . Comparison of adjusted attribuable risk estimator Stat Med 1992 11 2083 2091 . [PubMed: 1293670]
25. Gefeller O . Definitions of attributable risk--revisited Public Health Rev 1995 23 343 355 . [PubMed: 8638031]
26. Gefeller O , Land M , Eide GE . Averaging attributable fractions in the multifactorial situation: assumptions and interpretation J Clin Epidemiol 1998 51 437 451 . [PubMed: 9619972]
27. Greenland S . Bias in methods for deriving standardized mortality ratio and attributable fraction estimates Stat Med 1984 3 131 141 . [PubMed: 6463450]
28. Greenland S . Variance estimators for attributable fraction estimates, consistent in both large strata and sparse data Stat Med 1987 6 701 708 . [PubMed: 2825320]
29. Greenland S . Attributable fractions: bias from broad definition of exposure Epidemiology 2001 12 518 520 . [PubMed: 11505170]
30. Greenland S , Drescher K . Maximum-likelihood estimation of the attributable fraction from logistic models Biometrics 1993 49 865 872 . [PubMed: 8241375]
31. Greenland S , Robins JM . Conceptual problems in the definition and interpretation of attributable fractions Am J Epidemiol 1988 128 1185 1197 . [PubMed: 3057878]
32. Henderson BE , Ross RK , Pike MC . Toward the primary prevention of cancer Science 1991 254 1131 1138 . [PubMed: 1957166]
33. Kleinbaum DG, Kupper LL, Morgenstern H. Epidemiologic research: principles and quantitative methods. Lifetime Learning Publications, Belmont, 1982.
34. Kuritz SJ , Landis JR . Attributable risk estimation from matched-pairs case-control data Am J Epidemiol 1987 125 324 328 . [PubMed: 3812438]
35. Kuritz SJ , Landis JR . Summary attributable risk estimation from unmatched case-control data Stat Med 1988a 7 507 517 . [PubMed: 3368677]
36. Kuritz SJ , Landis JR . Attributable risk estimation from matched case-control data Biometrics 1988b 44 355 367 . [PubMed: 3390504]
37. Lagakos SW , Mosteller F . Assigned shares in compensation for radiation-related cancers (with discussion) Risk Anal 1986 6 345 380 . [PubMed: 3602505]
38. Land M , Gefeller O . A game-theoretic approach to partitioning attributable risks in epidemiology Biom J 1997 39 777 792 .
39. Land M , Vogel C , Gefeller O . Partitioning methods for multifactorial risk attribution Stat Methods Med Res 2001 10 217 230 . [PubMed: 11446149]
40. Last JM. A dictionary of epidemiology. Oxford University Press, New York, 1983.
41. Lrung HM , Kupper LL . Comparison of confidence intervals for attributable risk Biometrics 1981 37 293 302 . [PubMed: 7272416]
42. Levin ML . The occurrence of lung cancer in man Acta Unio Int contra Cancrum 1953 9 531 541 . [PubMed: 13124110]
43. Llorca J , Delgado-Rodriguez M . A new way to estimate the contribution of a risk factor in populations avoided nonadditivity J Clin Epidemiol 2004 57 479 483 . [PubMed: 15196618]
44. Lubin JH , Boice JR . Estimating Rn-induced lung cancer in the United States Health Phys 1989 57 417 427 . [PubMed: 2777548]
45. MacMahon B, Pugh TF. Epidemiology: principles and methods. Little, Brown and Co, Boston, 1970.
46. Msdigan MP , Ziegler RG , Benichou J , Byrne C , Hoover RN . Proportion of breast cancer cases in the United States explained by well-established risk factors J Natl Cancer Inst 1995 87 1681 1685 . [PubMed: 7473816]
47. Markush RE . Levin's attributable risk statistic for analytic studies and vital statistics Am J Epidemiol 1977 105 401 406 . [PubMed: 860704]
48. Mcelduff P , Attia J , Ewald B , Cockburn J , Heller R . Estimating the contribution of individual risk factors to disease in a person with more than one risk factor J Clin Epidemiol 2002 55 588 592 . [PubMed: 12063100]
49. Miettinen OS . Proportion of disease caused or prevented by a given exposure, trait or intervention Am J Epidemiol 1974 99 325 332 . [PubMed: 4825599]
50. Morgenstern H , Bursic ES . A method for using epidemiologic data to estimate the potential impact of an intervention on the health status of a target population J Commun Health 1982 7 292 309 . [PubMed: 7130448]
51. Ouellet BL , Romeder JM , Lance JM . Premature mortality attributable to smoking and hazardous drinking in Canada Am J Epidemiol 1979 109 451 463 . [PubMed: 443243]
52. Ries LAG, Eisner MP, Kosary CL, Hankey BF, Miller BA, et al. SEER Cancer Statistics Review, 1975–2001, National Cancer Institute, Bethesda, MD, 2004 (http://seer​.cancer.gov/csr/1975_2001). Accessed May 21, 2004.
53. Robins JM , Greenland S . Estimability and estimation of excess and etiologic fractions Stat Med 1989 8 845 859 . [PubMed: 2772444]
54. Robins JM , Greenland S . Estimability and estimation of expected years of life lost due to hazardous exposures Stat Med 1991 8 845 859 . [PubMed: 2006358]
55. Rothman KJ, Greenland S. Modern epidemiology. Lippincott-Raven, Philadelphia, 1998.
56. Schlesselman JJ. Case-control studies. Design, conduct and analysis. Oxford University Press, New York, 1982.
57. Seiler FA . Attributable risk, probability of causation, assigned shares, and uncertainty Environ Int 1986 12 635 641 .
58. Seiler FA , Scott BR . Mixture of toxic agents and attribuable risk calculations Risk Anal 1987 7 81 89 . [PubMed: 3615995]
59. Smith L. Person-years of life lost. In: Encyclopedia of biostatistics. Armitage P, Colton T., eds, Wiley, Chichester, 1998: 3324–3325.
60. Tuyns AJ , Pequignot G , Jensen OM . Le cancer de l'œsophage en Ille-et-Vilaine en fonction des niveaux de consommation d'alcool et de tabac Bull Cancer 1977 64 45 60 . [PubMed: 861389]
61. Uter W , Pfahlberg A . The concept of attributable risk in epidemiological practice Biom J 1999 41 985 999 .
62. Wacholder S , Benichou J , Heineman EF , Hartge P , Hoover RN . Attributable risk: advantages of a broad definition of exposure Am J Epidemiol 1994 140 303 309 . [PubMed: 8059765]
63. Wahrendorf J . An estimate of the proportion of colo-rectal and stomach cancers which might be prevented by certain changes in dietary habits Int J Cancer 1987 40 625 628 . [PubMed: 3679590]
64. Walter SD . The distribution of Levin's measure of attributable risk Biometrika 1975 62 371 374 .
65. Walter SD . The estimation and interpretation of attributable risk in health research Biometrics 1976 32 829 849 . [PubMed: 1009228]
66. Walter SD . Prevention for multifactorial diseases Am J Epidemiol 1980 112 409 416 . [PubMed: 7424889]
67. Walter SD . Effects of interaction, confounding and observational error on attributable risk estimation Am J Epidemiol 1983 117 598 604 . [PubMed: 6846317]
68. Walter SD . Attribuable risk in practice Am J Epidemiol 1998 148 411 413 . [PubMed: 9737552]
69. Whittemore AS . Statistical methods for estimating attributable risk from retrospective data Stat Med 1982 1 229 243 . [PubMed: 7187096]
70. Whittemore AS . Estimating attributable risk from case-control studies Am J Epidemiol 1983 117 76 85 . [PubMed: 6823955]