Jump to ContentJump to Main Navigation
Critical Appraisal of Epidemiological Studies and Clinical Trials$

Mark Elwood

Print publication date: 2007

Print ISBN-13: 9780198529552

Published to Oxford Scholarship Online: September 2009

DOI: 10.1093/acprof:oso/9780198529552.001.0001

Critical appraisal of a large population-based case–control study

Chapter:
(p.471) Chapter 15 Critical appraisal of a large population-based case–control study
Source:
Critical Appraisal of Epidemiological Studies and Clinical Trials
Author(s):

J. Mark Elwood

Publisher:
Oxford University Press
DOI:10.1093/acprof:oso/9780198529552.003.15

Abstract and Keywords

This chapter presents an example of the application of the scheme for critical appraisal: large population-based case-control study entitled ‘Risk of breast cancer in relation to lifetime alcohol consumption’, published in the Journal of the National Cancer Institute in 1995. This large case-control study has shown a regular positive association between recorded alcohol intake and breast cancer risk. The results are generally consistent with the results of other observational studies, most of which, however, would be open to the same limitations.

Keywords:   confounding, causal relationships, breast cancer, alcohol consumption

We will now discuss a large population-based case–control study, published in the Journal of the National Cancer Institute, June 21 1995, 87, 923–929 [1], and available at http://jncicancerspectrum.oxfordjournals.org. The abstract is reproduced here, courtesy of Oxford University Press and the first author.

Risk of breast cancer in relation to lifetime alcohol consumption

Matthew P Longnecker, Polly A Newcomb, Robert Mittendorf, E Robert Greenberg, Richard W Clapp, Gregory F Bogdan, John Baron, Brian MacMahon, Walter C Willett

Abstract

Background. Although an association between alcohol consumption and risk of breast cancer has been observed in many studies, questions of major importance remain, including the nature of the dose-response relationship and the effects of drinking at various periods in life.

Purpose. Our goal was to address the issues listed above with a large casestudy.

Methods. We conducted a population-based case–control study in Maine, Massachusetts (excluding the four counties that include metropolitan Boston), New Hampshire, and Wisconsin. Case patients were eligible if their diagnosis of invasive breast cancer was first reported to one of the four statewide cancer registries during the period of 1988 through 1991. During the accrual period, 11 879 potentially eligible case patients and 16 217 control subjects were identified. After excluding ineligible women from the study, telephone interviews were obtained from 6888 case patients and 9424 control subjects. Complete data for recent alcohol consumption, and thus final eligibility for study participation, were determined for 6662 case patients and 9163 control subjects. The average age at time (p.472) of interview was 58.7 years. The questions on alcohol use addressed average consumption during five periods of the subjects’ lives: ages 16–19, 20–29, 30–39, 40–59, and 60–74 years. Similar responses from 211 control subjects upon reinterview 6–12 months later were taken to be indicative of the reliability of the questionnaire used in this study.

Results. Lifetime average alcohol consumption (measured as the average grams per day consumed from age 16 to the recent past) and recent alcohol consumption (average grams per day consumed in the previous age interval) were associated with risk of developing breast cancer. The multivariate relative risk of breast cancer, in those who drink compared with abstainers, associated with average lifetime consumption of 12–18 g/day of alcohol (about one drink) was 1.39 (95% confidence interval [CI] = 1.16–1.67), of 19–32 g/day (about two drinks) was 1.69 (95% CI = 1.36–2.10), of 33–45 g/day (about three drinks) was 2.30 (95% CI = 1.51–3.51), and of greater than or equal to 46 g/day (four or more drinks) was 1.75 (95% CI = 1.65–2.64) (P for trend <.0001). The multivariate relative risk per 13 g/day (about one drink) of alcohol consumed before 30 years of age was 1.09 (95% CI = 0.95–1.24), whereas the relative risk associated with recent consumption of 13 g/day was 1.21 (95% CI = 1.09–1.34).

Conclusions. In these data, alcohol consumption was clearly related to breast cancer risk. Risk appeared to increase even at moderate levels of consumption. For women of all ages combined, consumption before 30 years of age was not an important determinant of risk.

A. Description of the evidence

  1. 1. What was the exposure or intervention?

  2. 2. What was the outcome?

  3. 3. What was the study design?

  4. 4. What was the study population?

  5. 5. What was the main result?

The exposure in this study is previous alcohol consumption, and the outcome is breast cancer; this is a case–control study. The association between alcohol consumption and breast cancer was already well established [2], as noted in the introduction, and so the objective of this study was to measure the association more precisely, particularly with regard to the dose-response relationship and the effects of alcohol consumption at different ages. Cases were obtained from population-based cancer registries, and controls from the corresponding general populations. The study population was women resident in four defined areas of the USA, who had cancer diagnosed and reported to the (p.473) corresponding registries between 1988 and 1991, and were aged under 75 years. Controls were selected from the general population within that age range from the same four areas. A prime issue in any case–control study is finding an appropriate listing of population members from which controls can be chosen. In this study the listings chosen were the drivers’ licence lists held at state level for subjects aged under 65, and the list of Medicare beneficiaries for subjects aged between 65 and 74. The interviews were carried out by telephone. For these reasons, the controls, if aged under 65, had to have a driving licence and a listed telephone number, and therefore similar restrictions were made for the cases. For the age range 65–74, all subjects should be on Medicare files, and therefore the restriction was only in terms of telephone listings. The full methods section as published is detailed and worth reading.

The main result was a positive association between increasing risk of breast cancer and increasing lifetime alcohol intake. Ex. 15.1 shows the results for lifetime average daily alcohol consumption, showing the ability of this large study to explore the dose relationship. Average daily alcohol consumption was divided into seven categories, and relative risks were calculated for each category with reference to the zero consumption category. The numbers of case and control subjects in each group, and the crude odds ratios (which in this study are equivalent to relative risks) are shown; these are calculated from the raw data by the methods shown in Chapter 3 and Appendix Table 1. These crude relative risks are not presented in the published paper as they do not take any of the confounding factors into account, but calculating the crude risks when appraising a paper is useful in understanding the data and may illustrate any major discrepancies. The published results show the relative risks adjusted for age and state by the Mantel–Haenszel method, as these are the chief demographic variables on which comparability is required. The ‘multivariate adjusted’ relative risks, adjusted for the range of variables described in the methods section, are also shown; these will be discussed under ‘confounding’ below. These three sets of results are generally quite similar. The most appropriate results to use for interpretation are the multivariate adjusted risk ratios, and these are given in the summary. There is a steadily increasing risk with increasing alcohol consumption except in the highest category, where the relative risk is lower than would be expected from a simple linear trend. The association can also be expressed as the relative risk per 13 g/day intake; this unit is used because it is equivalent to one drink of most types of alcoholic beverage. This is a useful parameter for making comparisons between subgroups, as will be shown. The statistical tests will be discussed in due course.

(p.474)

Critical appraisal of a large population-based case–control study

Ex. 15.1. Format of the results of the large case–control study showing the ability to explore dose-response relationships and confounder control. The crude odds ratios (not in the published table) are calculated directly from the data given; the published data were the relative risks (odds ratios) after adjustment for age and state only, and after adjustment for several confounders using a logistic regression model (see text). Data from Table 4 of Longnecker et al. [1]

B. Internal validity: consideration of non-causal explanations

6. Are the results likely to be affected by observation bias?

The hypothesis is that breast cancer occurrence is increased by alcohol intake in the years prior to that occurrence. Therefore the issue of observation bias concerns the relationship of the recorded information on alcohol intake to the biologically relevant level of alcohol intake. Non-differential error will reduce the observed relative risk towards the null svalue. Bias, i.e. a different relationship between recorded and true relevant alcohol use in the case compared with the control series, could have any type of effect on the result. The information was collected by a standardized telephone interview taking on average less than 25 minutes. The questions were quite detailed, addressing five different time periods, and consumption of beer, wine, and liquor separately. A typical question was ‘On average, how often did you drink one bottle, glass, or can of beer when you were in your twenties?’ The questionnaire instrument also included questions on lactation, hormone use, physical activity in adolescence and early adulthood, vitamin A intake, established breast cancer risk factors, and other characteristics. The data collected related to the time period prior to (p.475) diagnosis or to the ‘reference date’ for control subjects, defined as the date of interview minus the median time between diagnosis and interview for case patients in that state. This is to produce the same time interval between interview and the last relevant time of alcohol consumption for the controls as for the cases.

The main protection against bias is the standardization of methods, and so the questions were identical, and were presented in an identical fashion to both cases and controls. An important protection was that the interviewers were blind to the status of the interviewee at the start of the interview. Of course, many subjects would directly or indirectly make it clear to the interviewer whether they were breast cancer patients or not. The interview began with a request that participants not discuss their medical history until the end of the interview, and the interviewers were asked to record if they were still unaware of the case or control status of the interviewee by the end of the interview; this was so for 74 per cent of cases and 90 per cent of controls.

To assess the reproducibility of the questionnaire, 211 control subjects were re-interviewed after an interval of 6–12 months. The rank correlation coefficients between the average amount of alcohol consumed daily reported in the two interviews are presented for four different time periods; these range from 0.75 to 0.84. This re-test reliability is quite reasonable, although even a correlation of 0.8 suggests that an observed relative risk of 2.0 will relate to a true odds ratio of 2.4, based on the formula described in Chapter 5. However, this correlation of 0.8 is high compared with similar assessments of questionnaires or of clinical history items.

The critical issue is whether there could be bias. In the absence of a true difference, is there any reason why women who have had breast cancer would tend to report higher, or lower, alcohol consumption than comparison women from the general population? The authors note two studies which looked at re-test consistency of alcohol consumption assessed for both breast cancer cases and controls; they showed similar results for the two groups, suggesting that bias is unlikely. A useful extension of the study would have been to assess re-test reliability on a sample of cases as well as on controls. The literature reviewed in this publication suggests that reporting is generally reasonable, although some under-reporting is more likely amongst those who have very high alcohol intake. Such bias could affect the dose-response results seen in Ex. 15.1; assuming there is a true relationship, if cases with the highest alcohol consumption tended to under-report this, the risks in the highest categories may be underestimated.

Some of the interviews were carried out a considerable time after diagnosis, and it could be argued that differential recall bias might be particularly strong either a considerable time after diagnosis, or (perhaps more likely) shortly (p.476) after diagnosis. Results are given for cases interviewed within 14 months of diagnosis, and these were not substantially different from those based on all case patients. An analysis restricted to subjects where the interviewer was still blind to their status at the end of their interview would also be helpful, but is not reported. Thus some non-differential misclassification is inevitable, and so the reported relative risks are likely to underestimate the true relationship. Observation bias cannot be totally excluded, but it seems unlikely that the bias would be substantial enough to account for the main result, or to make a very large difference to it.

7. Are the results likely to be affected by confounding?

The issue of confounding is complex. There are a large number of factors which are known to alter the risk of breast cancer, including age, ethnic origin, social class, many aspects of reproductive history, diet, and obesity. A large number of factors would be expected to differ between women with different levels of alcohol consumption, including many of these same factors. Therefore the potential for confounding is very considerable. Individual matching for more than the general demographic factors would be difficult. Some matching on a frequency basis was used to make the ratio of controls to case subjects at least one in each 5-year group, and the groups were frequency matched by state. The main method of confounder control was to obtain information on other established risk factors for breast cancer, and use this in a multivariate analysis. Therefore the results are expressed in two ways. Relative risks adjusted simply for age in 5-year categories and state, by the Mantel–Haenszel method, are given, and in addition multivariate adjusted relative risks are shown. These relative risks are from a logistic regression model (of the type described in Chapter 6) with 6390 case patients and 8794 control subjects. As well as alcohol consumption, terms were included for age, state, age at first term pregnancy, parity, body mass index (an index of obesity), age at menarche, education, previous history of benign breast disease, and family history of breast cancer. The number of categories of each of these is given in the methods section. The use of oral contraceptives and replacement oestrogens were also considered as potential confounding factors, but in these data they were not correlated with alcohol consumption and therefore were not retained. As 98 per cent of the cases were white, race is not further considered. This multivariate adjustment did not make much difference to the results as shown in Ex 15.1; the multivariate adjusted relative risks are not greatly different from the results adjusted only for age and state.

There are two main issues with regard to confounding. One is whether the potential confounding factors that have been identified have been adequately dealt with. The ability to control their confounding depends on obtaining (p.477) accurate information on them, and using this information in an appropriate multivariate model. If the data on confounders have a high degree of error, this could compromise the ability to deal with confounding. While it is impossible to conclude that all confounding by these known factors has been totally removed, given that the approach used did not produce major differences in the association of alcohol intake and breast cancer, it is unlikely that further attempts to improve on confounder control by these variables would produce any greater difference.

The second issue is whether there could be confounding by other factors that have not been studied. This cannot be excluded. We need to consider whether any established risk factors for breast cancer, which could also be related to alcohol consumption, have not been considered in this study. One possibility is diet. No dietary information was collected, although aspects of diet have been considered as risk factors for breast cancer and could well be associated with alcohol consumption. However, the relationship of breast cancer was not very clear when this study was done; for example, a major meta-analysis of cohort studies showed no association with fat intake [3]. The evidence for an association with obesity, for example with weight gain in middle life, has become stronger since [4]. The associations seen with aspects of diet have not in general been stronger than that seen here with alcohol consumption; it is possible that the relationship with diet is confounded by alcohol consumption, rather than the other way round. There could also be some further as yet unrecognized factors, which would be important if they were true confounders, being a risk factor for breast cancer and also being independently association with alcohol consumption. There will also be factors which are intermediates in the association between alcohol consumption and breast cancer; indeed, it is very unlikely that the association, if causal, is direct in biochemical terms, and alcohol consumption may produce some metabolic or hormonal change which in turn increases the breast cancer risk. Such a factor would not be a confounder, as it would be an intermediate on the causal pathway, as discussed in Chapter 6.

8. Are the results likely to be affected by chance variation?

Chance variation is appropriately assessed by looking at the results after multivariate adjustment, as in Ex. 15.1. For all women, the relative risk for lifetime average consumption rises from the referent value of 1 in those with no alcohol consumption up to 1.75 in the highest category, and the values in all six categories used are individually statistically significant, with the lower 95 per cent confidence limit being greater than 1 (Table 4 in the original paper). (p.478) An appropriate overall statistical test is a test for trend; this is highly significant, with P < 0.0001. The statistical model used the square root of the alcohol consumption as this gave a better fit to the data, as described in the methods section. The overall association can also be expressed as the relative risk per 13 g/day, which is 1.31, with 95 per cent confidence limits of 1.20 to 1.43. The same format and statistical tests are applied to the results for more specific groups, as will be discussed. One of the objectives of this study was to compare the effects of alcohol consumption at different ages, as this should elucidate the mechanism of the association. We will discuss this further under the heading of specificity.

Summary: non-causal explanations

To summarize, observation bias has been as well controlled as is feasible in a retrospective study, but the possibility of biased responses from the subjects cannot be totally excluded. There will be some non-differential misclassification, although the re-test reliability of the methods used has been shown to be high. Confounding by most potential confounding factors has been adequately controlled by multivariate methods, but there remains the possibility of confounding by other factors not included in the study. Chance variation can be confidently excluded because the size of the study gives narrow confidence limits for the estimates of effect, and there is a regular dose-response relationship.

C. Internal validity: consideration of positive features of causation

9. Is there a correct time relationship?

Alcohol consumption after the recognition of cancer is clearly irrelevant to the causal hypothesis. Case patients were interviewed at a median of 14 months after diagnosis, and the interviews were restricted to asking about intakes up to the time of diagnosis, or a corresponding date for the controls. The accuracy of this cut-off might be less good where the interval from diagnosis to interview was longer, and no detailed information (e.g. the maximum time interval) is given. However, as noted before, the results were similar in those interviewed within 14 months of diagnosis to those in all subjects. The relevant time of alcohol intake depends on whether the relevant process is the first initiation phase of cancer development, or a later stage such as promotion, and so could be from a few to many years before clinical diagnosis.

10. Is the association strong?

The relationship is only moderately strong. For all women, the relative risks reach approximately 2.0 at the highest alcohol intake level measured. (p.479) Given that the range of alcohol consumption measured is quite wide, zero to approximately three to four alcoholic drinks per day, this size of relative risk is not particularly impressive. It is consistent with causality, and could arise if alcohol is a relatively minor contributor to breast cancer risk compared with other factors. Because of the high prevalence of alcohol consumption in the population, a relative risk of 2 or even less is important in clinical and public health terms. However, it does raise the question of whether alcohol intake is merely a rather inaccurate estimator of a more fundamental causal factor which, if identified, might show a much stronger association.

11. Is there a dose-response relationship?

The dose-response relationship is one of the main features of this study, as average daily consumption can be categorized over a considerable range. The dose-response relationship is shown for lifetime consumption in Ex 15.1, and data are also given for recent consumption, for consumption before age 30, and for pre-menopausal and post-menopausal women separately. In each of these groups there is a regular positive dose-response relationship. The study is large enough for the relative risks at moderate levels of consumption to have narrow confidence limits, and therefore they can be interpreted with some assurance. One of the objectives was to assess the risks related to moderate alcohol consumption, and so these results are given in the abstract; for example, for consumption of about one alcoholic drink per day, the relative risk is 1.39 with a 95 per cent confidence level of 1.16 to 1.67, a quite precise measure of effect. The regular dose-response relationships give considerable protection against the association being due to observation bias, as observation bias would be unlikely to be consistently related to the amount of alcohol consumption; we might expect observation bias to apply particularly to those with very heavy consumption, giving a rather erratic dose-response relationship. It is less protection against confounding, as if a confounding factor were regularly and systematically related to alcohol consumption, it could be consistent with a regular dose-response relationship.

12. Are the results consistent within the study?

and

13. Is there any specificity within the study?

In this study, specificity and consistency are assessed with respect to three factors: the consumption of alcohol at different ages, the occurrence of breast cancer at different ages, and the consumption of different types of alcohol.

(p.480) For consumption at different ages, detailed data are given in the paper for consumption during a recent age interval (but still preceding that in which diagnosis occurred) compared with consumption before age 30. The full results have been condensed in Ex. 15.2 to show only the overall relative risk per 13 g/day alcohol consumption, derived from the multivariate model. The analysis has to be restricted to subjects over age 40, as obviously this distinction cannot be made for younger subjects. There was a significant positive association with consumption during the recent time period (RR = 1.21), but the association with consumption before age 30 was weak and not significant (RR = 1.09). The data were also subdivided by age of diagnosis of breast cancer, into pre-menopausal and post-menopausal women, and the results then become more complex; the situation just described applied to post-menopausal women, but for pre-menopausal women the association with alcohol intake before age 30 was stronger than that for recent consumption, which is non-significant. Of course, these two time periods will overlap for many of these younger women, and so the interpretation is not as clear as it is for post-menopausal women.

The other comparison made, for women aged over 40, was to assess the joint effect of consumption before age 30 and in a recent time period. The results are given in the paper for 16 subgroups representing four categories of consumption in each time period, and the summary effects are shown

Critical appraisal of a large population-based case–control study

Ex. 15.2. Relative risks per 13 g/day alcohol consumption, from logistic regression model, in various subgroups. Data from Longnecker et al. [1]

(p.481) in Ex. 15.3. These results are ‘mutually adjusted’; that is the relative risk for alcohol consumption before age 30 is adjusted for alcohol consumption in the recent time period, and vice versa. The results show that there is no association with consumption before age 30 in women with high alcohol consumption in a recent time period, and vice versa. (Ex. 15.3, subtable A, relative risk 1.00; subtable B, relative risk 1.01). However, in women with no recent consumption, alcohol consumption before age 30 has a strong effect (RR = 1.72); similarly in women with no consumption before age 30, recent alcohol consumption has a strong effect (RR = 1.90). The authors suggest that this is consistent with a cumulative effect. These data do not support the prior hypothesis of a specific effect of alcohol consumption in early life.

With regard to the age of the women at diagnosis, comparing pre-menopausal with post-menopausal breast cancer showed that results were generally similar for these two groups. Although the association was stronger in post-menopausal than in pre-menopausal women, a statistical test of this interaction showed that the difference in effect was not significant.

Data are presented for daily intake of alcohol from three groups of drinks: beer, wine, and liquor. Using data on alcohol consumption in the most recent

Critical appraisal of a large population-based case–control study

Ex. 15.3. Relative risks per 13 g/day alcohol consumption, from logistic regression model, by age at consumption, stratified for consumption at other ages. Data from Table 3 of Longnecker et al. [1]

(p.482) age interval, the multivariate model was fitted with a continuous variable for each measure of alcohol consumption. As these factors are fitted simultaneously, the resulting relative risk is the estimate of the change in risk with a unit change in alcohol consumption, here categorized as 13 g/day, adjusting for the other sources of alcohol. The relative risk estimates for beer, liquor, and wine were 1.25, 1.18, and 0.93, respectively; the coefficients for beer and liquor are statistically significant, but that for wine is not significant. Thus the results show no association with wine, but positive associations with beer and liquor. We need to assess this result with regard to possible observational bias and confounding. The weaker association with the consumption of wine could arise if the degree of non-differential classification for this exposure were greater than that for the other types of alcohol, or if there were other confounding effects related to this exposure. If there is under-reporting of heavy consumption, this reduces the range of reported consumption, which thus reduces the ability to detect a real association; such an effect might have a greater influence where the range may be relatively small, as in wine consumption.

Conclusions with regard to internal validity

The internal validity of the result of this study is high. It is a large study. It has been carefully conducted with standardized and, to a large extent, interviewer-blind data collection, although of course the subjects providing the data are well aware of whether or not they have had breast cancer. The possibility of observation bias remains, but the regularity of the dose-response relationships, the good result from an examination of re-test reliability of the questionnaire, and previous information suggesting that interview data collected in this way are reasonably reliable and unlikely to be influenced by bias, together make observation bias a less likely explanation of the results. Non-differential error will make the results underestimate the true association. Confounding by the accepted risk factors for breast cancer has been dealt with by collecting data on these items and controlling them in a multivariate analysis; the fact that this gave very little change in the risk ratios suggest that confounding by these factors is minor, and also suggests that further control, for example by having more comprehensive data on such factors, would be unlikely to greatly modify the result. However, the influence of unrecognized confounders cannot be dismissed. Diet has not been assessed and could be a confounder. Chance variation can be much more confidently excluded, as the study is large, and the tests show that the probability of such results occurring by chance is extremely small. The time relationships are consistent with causality, the relationship is reasonably strong, and there is a very clear and regular dose-response relationship. The results suggest generally similar affects for (p.483) pre-menopausal and post-menopausal women, and suggest that the relationship with alcohol consumption is confined to alcohol taken in the form of beer or liquor, rather than wine. A causal explanation of the association seems more likely than any of the alternative non-causal explanations, and it is reasonable to proceed on this basis. The most plausible non-causal explanation would be confounding by other dietary factors, or other unrecognized factors.

D. External validity: generalization of the results

14. Can the study results to applied to the eligible population?

The relationship of participants to eligible subjects brings us to the question of response rates and missing data. Ex. 15.4 shows the typical complexity of a large population-based case–control study. The 6862 cases participating fully in the trial are derived from an eligible population of 8579. Of these, 14.9 per cent were not approached for interview for the reasons shown, 5.6 per cent of those approached refused the interview, and a further 3.3 per cent had missing data on some of the key variables. Therefore the voluntary response rate is extremely high at 91 per cent, and the participation rate (the participant group divided by the eligible population) is also substantial at 78 per cent.

For the control group, the 9163 controls are derived from 11 238 eligible control subjects. The proportion not approached (2.5 per cent) is much lower than that of the cases, because fewer died before interview and there was no requirement for a doctor to give permission. On the other hand, the 14 per cent refusal rate for interviews is much higher than that of the cases, although it is still very reasonable; a similar proportion of the interviews (2.8 per cent) were incomplete. The voluntary response rate for controls was 83.6 per cent, which is very high for a study of this nature in the USA, and the participation rate at 81.5 per cent is also high and very similar to that of the case series.

These results are impressive. The cases will tend to exclude women with very advanced disease who died before interview, and those whose doctor refused permission may also have had more advanced disease. This is relevant if the association between alcohol and breast cancer also relates to the extent of disease at diagnosis, for example by affecting the growth speed of the tumour or the speed of diagnosis. Apart from that, the exclusion and voluntary refusals are minor. We might well expect that they could be related to alcohol consumption, and that the higher refusal rate amongst those approached for interview in the control group compared with the cases could lead to the exclusion of more controls with particularly high alcohol intakes. If this bias did occur, it would tend to exaggerate the observed relative risks; the recorded level of alcohol (p.484)

Critical appraisal of a large population-based case–control study

Ex. 15.4. Derivation of the case and control groups contributing to the analysis of alcohol consumption and breast cancer in Longnecker et al. [1]

consumption in participating controls would underestimate the consumption in all eligible controls, with this bias occurring to a lesser extent among the cases.

15. Can the study results be applied to the source population?

The definition of the source population is, for the cases, women diagnosed with breast cancer who were resident in the four geographical areas in 1988–1991, and for the controls, women in the general population for the four areas. The eligibility criteria that restrict the study and were applied to both cases and controls were that the women had to have a listed telephone number (which excluded 22 per cent of cases and 27 per cent of controls) and possess a drivers’ licence (which excluded 3 per cent of cases and 1 per cent of controls) (Ex. 15.4). There is (p.485) no discussion of the effects of these exclusions. There should be little effect on internal validity, as the restrictions apply to both cases and controls, and an internal bias will only occur if the selection criteria had differential effects for cases compared with controls. The restriction by the telephone criterion may mean that women in unfavourable socio-economic circumstances may be under-represented; this restriction might exclude some women who have alcohol-related problems. Some women with particular occupational roles might be excluded, given that many telephone numbers in the USA are unlisted. However, it seems unlikely that these exclusions would greatly affect the relationship between alcohol consumption and breast cancer occurrence examined in this study.

16. Can the study results be applied to other relevant populations?

This study shows an association in recently diagnosed women in the USA, who were almost all white. Breast cancer is a disease that varies considerably in frequency, being more common in populations of European origin than in, for example, Asian or African populations. Alcohol consumption in women is a culturally specific exposure, and its associations with other social, economic, and cultural factors vary considerably in different societies. This means that the confounding relationships could be different in different societies. On the other hand, despite the considerable range of incidence rates of breast cancer, other breast cancer risk factors such as reproductive factors have reasonably consistent associations with the disease in a wide range of cultures. A fundamental biological association between alcohol consumption and breast cancer would be expected to be universal. Therefore, in general, the reasonably high internal validity suggests a biological causal relationship between alcohol consumption and breast cancer, which can be applied widely. The details of the relationship, such as the specificity to consumption of beer or liquor rather than wine, might suggest possible differences in confounding relationships, in that the social factors linked to these different drinks may vary. It would be reasonable to consider these results as applicable to generally similar societies, i.e. affluent Western societies with women of predominantly European origin, but there would be more caution in applying the results to other social, ethnic, or racial groups.

E. Comparison of the results with other evidence

17. Are the results consistent with other evidence, particularly evidence from studies of similar or more powerful study design?

When published, this study, primarily because of its size, was one of the best individual case–control studies of this topic. Other case–control (p.486) and cohort studies are reviewed in the discussion. Many studies have found a positive association between alcohol consumption and breast cancer; the inconsistencies are in terms of whether drinking at different ages has different affects, and in terms of specific relationships to the type of alcohol. The further studies published after this one are briefly discussed later in this chapter.

18. Does the total evidence suggest any specificity?

The main issues with regard to specificity on the basis of the total evidence are the two which have been discussed, age specificity in the exposure, and specificity to different types of alcohol. There are no consistent results from the total range of human-based evidence available for either of these aspects, and therefore no firm conclusion can be reached.

19. Are the results plausible in terms of a biological mechanism?

The possible biological mechanisms of this association are discussed in some detail in this paper. Several mechanisms by which alcohol intake could increase breast cancer risk have been suggested, although none of these is established. Some animal experiments have shown that alcohol intake increases the risk of breast tumours and the proliferation rate of breast tissue cells, although the animal results are also inconsistent. Alcohol is not generally accepted as a simple carcinogen, i.e. an initiator of the cancer transformation process in cells. However, it is recognized as a promoter, i.e. a chemical which promotes the development of cells which have already gone through the first cancer transformation step, by acting on other aspects of cancer cell changes, or through hormonal, immunological, or other mechanisms. Therefore we can conclude that the association is plausible, although the mechanism is unclear. The authors quote studies showing that alcohol appears to increase serum oestradiol levels, which is relevant as there is other work linking high oestradiol levels to breast cancer risk.

20. If a major effect is shown, is it coherent with the distribution of the exposure and the outcome?

The overall association is not strong enough to make an argument of coherence tenable. Breast cancer is a common disease, and several other factors have associations as strong as or stronger than that produced here. Therefore, in general, we would not anticipate a clear relationship between alcohol consumption and breast cancer incidence, for example on a geographical or secular trend basis. Having said that, the high incidence rate in Western countries, and perhaps the increasing trend in breast cancer incidence over recent (p.487) decades, is consistent with differences in alcohol consumption. A positive association between breast cancer mortality and estimated alcohol intake per capita, comparing countries, has been shown [5], although the association with fat intake was greater. The limitation is that this type of evidence is open to so many other interpretations that it adds little to the argument for causality; the current study uses a much more powerful method of inquiry.

Conclusions

In conclusion, this large case–control study has shown a regular positive association between recorded alcohol intake and breast cancer risk (Ex. 15.5). The most convincing aspect of this is the clear and regular dose-response relationship. The least convincing aspect is the complex relationship with consumption at different ages, with curious results when age at diagnosis is also taken into consideration, which do not lead to a clear conclusion. The overall association is unlikely to be due to observation bias. Confounding by most well-established risk factors for breast cancer has been dealt with. Confounding by diet has not been addressed, and the disease is sufficiently complex that the possibility of unrecognized confounding factors remains. The associations seen are highly statistically significant. The results are generally consistent with the results of other observational studies, most of which, however, would be open to the same limitations. A reasonable conclusion is that a causal explanation is the most likely explanation of the results seen, but that confounding cannot be excluded. The aspect of the results suggesting that the association is specific to beer or liquor consumption, rather than wine consumption, is not consistent with the totality of other evidence, and can be regarded only as tentative.

Subsequent development

This study was a substantial contribution to knowledge about alcohol consumption and breast cancer, and its most important result was to show that risk increased even at moderate levels of consumption, of even one standard drink per day. The other major finding was that consumption before the age of 30 was less important than average lifetime consumption or alcohol consumption in recent years. The association with even low levels of consumption has been confirmed, particularly in a large pooled analysis of 53 epidemiological studies, including over 58 000 women with breast cancer, done by the same group who explored other risk factors for breast cancer, such as oral contraceptive use and a previous abortion, as discussed earlier in this book [6]. This meta-analysis showed that the risk of breast cancer increased by 7 per cent for each extra unit of alcohol (10 g) consumed per day, and estimated (p.488)

Critical appraisal of a large population-based case–control study

Ex. 15.5. Summary of assessment of the case–control study of breast cancer and alcohol consumption: Longnecker et al. [1]

(p.489) that about 4 per cent of breast cancers in developed countries were attributable to alcohol use. Of course, the individual studies have given varied results; for example, the well-known Framingham cohort study did not show any relationship of alcohol to breast cancer, and reasons for this have been discussed [7]. The finding that risk is increased more by recent alcohol intake than by consumption in early life has also been shown in some other cohort and casestudies [8,9], but there is less consistency on this [10]. The issue of whether risk of breast cancer varies by the type of alcohol consumed has not been confirmed, with some other studies showing no such variation [10,11]. The association with alcohol use may also vary with the type of breast cancer, being greater with oestrogen-receptor-positive tumours [12]. Considerable work has been done on the possible mechanisms for the association of breast cancer with alcohol, which may be closely related to aspects of obesity and to folate consumption and metabolism [13,14]. Alcohol intake is associated with several cancers in addition to breast cancer; again, the mechanism of its carcinogenic effect is not yet known, and the possibilities include toxic effects of the main metabolite acetaldehyde, increased oestrogen concentration, the production of reactive oxygen and nitrogen compounds, a solvent action in combination with tobacco carcinogens, and effects on folate metabolism [15].

References

Bibliography references:

1. Longnecker MP, Newcomb PA, Mittendorf R, Greenberg ER, Clapp RW, Bogdan GF et al. Risk of breast cancer in relation to lifetime alcohol consumption. J Natl Cancer Inst 1995; 87: 923–929.

2. Longnecker MP. Alcohol consumption in relation to risk of breast cancer: meta-analysis and review. Cancer Causes Control 1994; 5: 73–82.

3. Hunter DJ, Spiegelman D, Adami H-O, et al. Cohort studies of fat intake and the risk of breast cancer: a pooled analysis. N Engl J Med 1996; 334: 356–361.

4. Holmes MD, Willett WC. Does diet affect breast cancer risk? Breast Cancer Res 2004; 6: 170–178.

5. Schatzkin A, Piantadosi S, Miccozzi M, Bartee D. Alcohol consumption and breast cancer: a cross-national correlation study. Int J Epidemiol 1989; 18: 28–31.

6. Hamajima N, Hirose K, Tajima K, et al. Alcohol, tobacco and breast cancer: collaborative reanalysis of individual data from 53 epidemiological studies, including 58,515 women with breast cancer and 95,067 women without the disease. Br J Cancer 2002; 87: 1234–1245.

7. Longnecker MP. Invited commentary: the Framingham results on alcohol and breast cancer. Am J Epidemiol 1999; 149: 102–104.

8. Horn-Ross PL, Canchola AJ, West DW, et al. Patterns of alcohol consumption and breast cancer risk in the California Teachers Study cohort. Cancer Epidemiol Biomarkers Prev 2004; 13: 405–411.

(p.490) 9. McDonald JA, Mandel MG, Marchbanks PA, et al. Alcohol exposure and breast cancer: results of the women’s contraceptive and reproductive experiences study. Cancer Epidemiol Biomarkers Prev 2004; 13: 2106–2116.

10. Singletary KW, Gapstur SM. Alcohol and breast cancer: review of epidemiologic and experimental evidence and potential mechanisms. JAMA 2001; 286: 2143–2151.

11. Petri AL, Tjonneland A, Gamborg M, et al. Alcohol intake, type of beverage, and risk of breast cancer in pre- and postmenopausal women. Alcohol Clin Exp Res 2004; 28: 1084–1090.

12. Suzuki R, Ye W, Rylander-Rudqvist T, Saji S, Colditz GA, Wolk A. Alcohol and post-menopausal breast cancer risk defined by estrogen and progesterone receptor status: a prospective cohort study. J Natl Cancer Inst 2005; 97: 1601–1608.

13. Dumitrescu RG, Shields PG. The etiology of alcohol-induced breast cancer. Alcohol 2005; 35: 213–225.

14. Le Marchand L, Haiman CA, Wilkens LR, Kolonel LN, Henderson BE. MTHFR polymorphisms, diet, HRT, and breast cancer risk: the multiethnic cohort study. Cancer Epidemiol Biomarkers Prev 2004; 13: 2071–2077.

15. Boffetta P, Hashibe M. Alcohol and cancer. Lancet Oncol 2006; 7: 149–156.