Vegan Society Briefing

Milk and Breast Cancer (3 of 7)

Stephen Walsh, November 2001.

Is there an association between milk and breast cancer?

How are associations between diet and disease established?

Before looking specifically at milk and breast cancer, it is useful to consider some general methods by which researchers have attempted to establish dietary causes for human cancers by studying people. Readers already familiar with nutritional epidemiology can skip this section, as it is largely standard material.

Regional or "ecological" comparison

A common form of study involves comparing variations in disease rates in different countries or regions with variations in diet. This will often show associations that are unlikely to be due simply to chance, but there are many factors that vary between countries and most apparent associations are not causal. It is particularly difficult to differentiate between the effects of the many distinct changes that occur with economic development. These typically include increased consumption of animal products and fat, adequate calorie intake throughout life, less physical activity, access to effective sanitation and medical services and having fewer children at a later age. For this reason, regional comparison is regarded as a weak form of evidence. A strength of inter-country comparisons is that characteristics such as average diet are well defined and vary over a wide range.

Case control studies

Another source of evidence is provided by case control studies. In such studies, a group of people with established disease ("cases") are compared with a group of people who are free from the disease ("controls"). Statistical analysis is then used to identify characteristics that differentiate the cases and the controls and the results are expressed as a relative risk (RR) for the disease associated with such characteristics. The results will usually be given in terms of the value of the relative risk comparing two groups. Typically the overall group will be divided into thirds (tertiles), quarters (quartiles) or fifths (quintiles) after ordering the group members from the lowest to highest values of some characteristic. The relative risk will compare the group with the highest values of that characteristic with the group with the lowest values, e.g. upper quintile (fifth of group with highest values for characteristic) vs lower quintile.

A relative risk of 1 means no difference in risk between the two groups. The probability that the relative risk differed from 1 purely by chance will also be reported. If the probability of a relative risk being different to 1 by chance is less than 0.05 that association is usually described as statistically "significant". The choice of 0.05 as a threshold is a compromise between getting an excessive number of false associations due to chance and missing a real relationship, and is to some extent arbitrary. For example, a relative risk of 1.5 between the upper quintile and the lower quintile with a probability of 0.003 would suggest a moderate and clearly statistically significant association, i.e. very unlikely to be due to chance. The RR of 1.5 means that the upper quintile show 50% greater risk of the disease being studied than the lower quintile. A relative risk of 5 with a probability of 0.04 would indicate a strong effect with modest statistical significance. A relative risk of 10 with a probability of 0.3 means little, as the observed relative risk could easily be a chance observation so the result is not significant.

There are two major weaknesses in case control studies looking at diet. Firstly, the controls may not be representative of the general population being studied (selection bias). Often controls show a greater interest in health than most people, making them more likely to volunteer, and they may therefore show an unusually high consumption of foods believed to be healthy. This can lead to a false conclusion that these "healthy" foods make development of the disease less likely, where in fact the controls were simply unrepresentative of the disease-free population. Secondly, cases' current diet may be altered as a consequence of the disease and their recollection of past diet may also be changed (recall bias). A simple example of recall bias is that people with hip fractures tend to underestimate past milk consumption. This underestimation makes case control studies likely to conclude that milk consumption is protective against fractures, regardless of the true effect. Because of these biases, case control studies are viewed with considerable scepticism. However, case control studies have the potential strength that, because relatively few people are involved, it is possible to carry out very thorough evaluations of diet. This can make them more likely to detect a true effect than studies that rely entirely on simple questionnaires to evaluate diet.

Prospective studies

Prospective studies overcome the biases of case control studies. In a prospective study a sample of the population (called a cohort) who are free from the diseases of interest at the start of the study is examined. As there are no cases at this stage, selection and recall biases do not come into play. As in case control studies, the volunteers may be healthier than the general population but so long as comparisons are made within the cohort this effect will not introduce a bias. Comparisons between the study cohort and the general population are, however, subject to selection bias. To provide increased confidence that pre-existing disease is not affecting the initial observations it is common practice to eliminate cases arising within a few years of the start of the study from the subsequent analysis. If a significant relative risk is observed in a prospective study that has followed these precautions then it is more likely to be a true association than one obtained from a case control study. A weakness of prospective studies is the difficulty of accurately characterising the population studied without excessive cost as they must include large numbers of people to ensure that a reasonable number of cases will arise during the study. Inaccurate characterisation may lead to true effects being missed.

Confounding

Ecological, case control and prospective studies are all subject to confounding - where an apparent effect from one characteristic is actually due to a different but related characteristic. As mentioned previously, this problem is particularly severe in comparisons between countries based on a characteristic which changes with economic development. Any such characteristic will be closely associated with many other characteristics, so cause and effect cannot be inferred. Once a behaviour is perceived to be healthy, new possibilities for confounding arise. If people who are at particular risk of a disease adopt a behaviour believed to be protective, then future studies may find the protective behaviour to be associated with increased risk. This may apply, for example, to small boned individuals and individuals with a family history of fracture increasing their calcium intake to reduce fracture risk.

If several characteristics are strongly associated, it can be difficult to separate out the causal characteristic even if all the potentially relevant characteristics are measured. If a relevant characteristic is not measured or not analysed, a false association may readily arise. A key example of this is that intake of most foods is strongly associated with total energy intake, which in turn may be associated with obesity, physical activity and hormone levels. It is therefore standard practice to use statistical methods to adjust for energy intake so that this effect is removed from the analysis. Age must also be adjusted for, as it is an important factor in every disease. Other known risk factors for a particular disease should be measured and adjusted for. Risk of breast cancer rises with earlier age of menarche, later age at first live birth, later age at menopause, increased BMI (weight in kg divided by height in metres squared) in postmenopausal women, and increased height. It would be normal to present results after adjustment for these factors. As some of these factors are themselves dependent on diet it can be useful to present results with adjustment for age and energy only, as well as presenting fully adjusted results. If the partially and fully adjusted risks are very different, particularly careful analysis is required to try to separate the influence of diet.

Randomised intervention studies

The gold standard in investigating diet would be to randomly assign individuals to different diets for many years and compare the outcome. This is generally not practical as people are unlikely to cooperate and large-scale trials are very expensive. However, it is possible to carry out relatively short term or partial dietary modifications. Results from such trials are unlikely to be due to confounding, but the duration may be too short to be definitive in relation to the lifetime effect of a behaviour. If the duration of the study is too short to evaluate the final outcome of interest, such trials may use intermediate risk markers as an outcome measure, e.g. change in bone mineral density or biochemical markers of bone resorption following calcium supplementation.

In practice, a combination of randomised intervention studies and prospective studies provides the best evidence. If short duration intervention studies and longer duration prospective studies show conflicting results, then it is difficult to draw firm conclusions.

The particular strengths and weaknesses of each method should be borne in mind in looking at results on diet and health.

Back | Next     

Page 1 | 2 | 3 | 4 | 5 | 6 | 7