Context of this post
In a previous post I have been critical of much current nutrition research where intake of specific foods or nutrients is said to be linked to increased or decreased risk of specific diseases and I promised to explore some of these “food claims” in future posts.
Origins of the hypothesis
Galactosaemia is an uncommon inherited condition in which those affected lack an enzyme necessary for the metabolism of the sugar galactose whose concentration in blood consequently rises (hence galactosaemia). Dietary galactose is derived principally from the breakdown of lactose or milk sugar although free galactose is found in some fruits and vegetables especially figs, dates, peas and beans. In yogurt some of the lactose has been broken down into glucose and galactose so yogurt has substantial amounts of “free” galactose.
It has been known for several decades that women with galactosaemia usually have impaired ovarian function leading to fertility problems and premature menopause. High galactose concentrations are thought to be toxic to ovarian cells thus hastening the normal menopausal decline in ovarian function. Studies with animals exposed to very high galactose diets (way beyond those found in any human diets) lends some support the idea that very high exposure to galactose impairs ovarian function.
On the strength of these observations it was hypothesised that in galactosaemia, reduced output of ovarian hormones would occur due to exposure of the ovaries to high galactose concentrations. This reduced ovarian function would lead to a compensatory increase in output of ovary-stimulating hormones from the pituitary gland as happens normally at the menopause. Overstimulation of the ovary by these pituitary hormones might increase the chances of promoting the development of ovarian cancer. Despite this suggestion there is no evidence that women affected by galactosaemia have any increased risk of developing ovarian cancer. This theory was widened to suggest that diets high in galactose (i.e. high in dairy products) might increase the risk of ovarian cancer in all women.
How the idea of a link between dairy products and ovarian cancer was popularised
A case-control study by Daniel Cramer and his colleagues published in the Lancet in 1989 did much to popularise the notion that consumption of certain dairy products might increase the risk of ovarian cancer. This study tried to assess the diets of around 240 cases (i.e. women with ovarian cancer) and disease-free controls in the years before diagnosis. They found no difference in lactose consumption between the two groups but when lactose was related to activity of an enzyme which metabolises it then this ratio was higher in the cancer cases. Several individual dairy foods were considered separately and there was a small but statistically significant positive association between both yogurt and cottage cheese consumption and ovarian cancer risk. This study was widely reported in the media e.g. a UPI newswire of July 12th 1989 by Rob Stein (Dairy products linked to ovarian cancer) and a piece in The New York Times dated July 25th 1989. The general tone of the press coverage would certainly have been alarming to some women who ate lots of these dairy products especially if they had had some family history of ovarian cancer. The UPI piece starts with the statement:
“A new study suggests that eating large amounts of dairy products, especially yogurt and cottage cheese may increase the risk of developing ovarian cancer…”
The lead author is quoted as saying that he:
“stressed the findings need to be confirmed before recommending women eat less dairy products”.
“the findings were cause for concern, especially for women who eat a lot of yogurt”.
As discussed later, this study and its media coverage helped to stimulate the publication of numerous other studies over the next 27 years looking at the relationship between consumption of dairy foods and ovarian cancer. Once a claim is published then other authors can use it to justify attempts to refute or support it and increase the chances of publishing their own findings.
Criticisms of the paper by Cramer and his colleagues
Case-control studies are observational (epidemiological) studies that are well down the hierarchy of evidence in medical science. They have played an important role in establishing some environmental/lifestyle factors as causes of ill health e.g. smoking and lung cancer, asbestos exposure and mesothelioma, front sleeping and cot death but in these examples the relative risk of exposure in the cases is many times that in the healthy control group e.g. well over ten times in the first two examples. People with lung cancer are more than ten times as likely to have smoked regularly as a matched group of healthy people.
I first read this Lancet paper of Cramer and his colleagues in detail because I was looking for an example of a case-control study to discuss in one of my books that illustrated some of the problems and limitations of case-control studies. In a previous post I have discussed the claim of Young and Karr (2011) that most claims made on the basis of observational studies are wrong. In this post, I reviewed some of the problems with such studies that make false claims more likely. Many of these flaws are found in this paper and some of them are summarised below.
- Crude dietary assessment and classification. Subjects were asked to fill in a questionnaire that sought information about the frequency and amounts of 116 food items consumed over the previous five years. Cases were asked to ignore changes in their diet since diagnosis; this is clearly an important point because otherwise one could be measuring the effect of ovarian cancer, cancer treatment or cancer awareness on the women’s diets rather than their diet as a potential cause of the cancer. How reliable are these past dietary estimates? What if the effects of early ovarian cancer affected dietary habits and preferences before their diagnosis? Ovarian cancer is often only diagnosed when it is fairly advanced. The galactose intake of the women was estimated from the amounts of 11 different dairy foods women said they ate on the “semi-quantitative” food frequency questionnaire; surely this can only be termed a rough estimate? For each of the 11 dairy products, the subjects were divided into two categories on the basis of whether they ate a specified amount of the food a) less than monthly or b) more than monthly e.g. whether they ate 226g (8oz) of yogurt less than or more than monthly. Once again a pretty crude classification.
- Poorly matched case and control groups and multiple modelling. The 235 cases of ovarian cancer were white women with primary ovarian cancer living in the Boston area who the team were able to contact and who agreed to participate. Using phone book listings they were matched to 239 control women on the basis of race, residential area and approximate age. When other characteristics of the two groups were compared, it was found that the cases were more likely to be Jewish, college educated, never married, never had a child and to have never used oral contraceptives. They were thus fairly well matched for age and race but differed in quite a number of other ways which could well have affected their susceptibility to ovarian cancer. It might well be that women with some such characteristics are both more likely to get ovarian cancer and more likely to eat different amounts of dairy products. They did attempt some statistical adjustment to allow for these differences but this is a very imperfect process and who is to say what other relevant differences there may have been between the groups that were not allowed for?
- Multiple analyses. The initial hypothesis was that high galactose intake from milk sugar (lactose) is causally linked to ovarian cancer. They found no significant difference between the total galactose intakes of cases and controls. They then looked at differences in the consumption of eleven dairy foods that contain galactose. Using the very crude categorisation described earlier and after “adjustment” for the differences between the groups they reported that cases were 1.7 times more likely to eat yogurt at least once a month and 1.4 times more likely to eat cottage cheese monthly; none of the other reported comparisons were significantly different. These two differences are just statistically significant but the more factors one tests, the more likely statistically significant false positive results become. I wonder how many of the other 116 foods in the questionnaire were consumed in significantly different amounts in the two groups. Although galactose consumption per se was not significantly different between the two groups, for just over half of the subjects they also measured the activity in venous blood of a transferase enzyme involved in galactose metabolism. It is a severe deficiency of this transferase enzyme that causes galactosaemia. The women were then subdivided into four categories: low lactose/high transferase activity; low/lactose/low transferase activity; high lactose/high transferase activity; high lactose low transferase activity. The relative risk of ovarian cancer in these four categories was 1.0, 1.1, 1.3 and 2.2 respectively. Whilst the relative risk in the last group might be just significantly different to that in the first category, how confident can one be that this is not just a statistical freak obtained when rather imprecise and flexible derived values are compared?
- Small effect size. Given the very crude and weak methodology which these authors were forced to use, then to demonstrate any convincing relationships one would need to be seeing substantial differences between the case and control groups. If they had found say five-fold differences between the case and control groups then one might believe that they had identified real potential causal factors that warranted more thorough investigation. Given the opportunities for error and bias inherent in these studies then I am inclined to regard any differences reported as probably just statistical flukes or measures of the bias in the group selection and methodology.
- Underpowered study. The number of willing study participants with confirmed primary ovarian cancer that they could recruit for a study based in one city is clearly limited. Are these 235 cases representative of women with ovarian cancer? Even more questionable is whether the 239 control women were a representative sample of women of the same age, race and social class who did not have ovarian cancer. There were clearly a number of quite important differences between the two groups that could have been and probably were determinants of ovarian cancer risk. For the enzyme studies the sample size was almost cut in half largely on the basis of which women were willing to contribute blood samples. The mean activity of the enzyme was lower in the cases than the controls. Might this be a direct or indirect effect of the cancer or its treatment? Given this, how much weight should one give to measures of lactose: transferase activity ratios?
- Publication bias. Publication bias describes the observation that positive and hypothesis-confirming results are more likely to be published than negative results (see here) and thus authors feel pressured to produce positive results. The authors and certainly the press coverage of this study emphasised the positive relationships, particularly the small increase in relative risk associated with eating yogurt or cottage cheese once a month. The results failed to support the primary hypothesis of higher galactose consumption in women prior to developing ovarian cancer. One could well imagine that a different set of authors might well have presented the results of this same study in a different way. For example, authors with affiliation to the dairy industry might well have concluded that it provides no evidence to support the suggestion that high galactose intake causes ovarian cancer. Even if they reported the apparently positive data relating to yogurt and cottage cheese they might well have framed it much more conservatively e.g. we cannot totally rule out the possibility that some dairy foods are very slightly increased in the case group but this is highly likely to be a chance observation resulting from the crudeness of the methodology and the multiple testing. They might have noted the reduced activity of the transferase enzyme as a possible direct or indirect effect of the cancer or treatment. I wonder how the referees and editorial board of the Lancet back in 1989 would have responded to this paper if it had been couched in these negative terms.
My criticisms of this paper are not meant to be a criticism of the scientists involved. They had put forward an interesting and credible hypothesis worthy of screening. They were limited by the inevitable fallibility of the methodology available to them e.g. it is notoriously and probably intractably difficult to accurately estimate human dietary intakes especially past intakes. The control group, who were selected in quite a reasonable way, had many characteristics that were different to the cases and the methods available for adjusting for such differences are inevitably crude. If they had found substantial differences in consumption patterns between the groups then this might have been a solid justification for further studies aimed at testing their hypothesis. The differences were actually small and unconvincing but in my opinion these differences were unreasonably stressed especially in the way the data was presented in the media.
Later studies testing this hypothesis
In the 27 years since the publication of this paper by Cramer et al, their hypothesis has spawned dozens of studies of different types:
- More case-control studies
- Many cohort studies where dairy food and lactose is recorded in large groups of women and related to the subsequent development of ovarian cancer in a few of them
- Studies in which the effects of high galactose intake is studied in animals
- Laboratory studies looking at the effect of exposure of ovarian cells to high galactose concentrations.
There have also been a number of attempts at aggregating the epidemiological studies (meta-analysis) to try and get a consensus of their findings. For example Jeanine Genkinger and 25 colleagues in 2006 reported the results obtained by pooling the results of 12 cohort studies involving a total of more than 550,000 women. When they estimated risk of women who consumed more than 30g/day of lactose and compared it to those who consumed less than 10g/day the relative risk was marginally higher (19%) which was just statistically significant but well within the range of being due to potential bias. Note that a pint of milk contains around 20g of lactose. Five of the individual studies found that relative risk was increased in the lower lactose consumers but seven found that it was greater in the high lactose consumers although in not one of the 12 individual studies was this difference statistically significant despite each involving tens of thousands of women. They found no associations with particular dairy foods like milk, cheese or yogurt and ovarian cancer.
The most recent of these meta-analyses that I have found was published in 2015 by Jing Liu and colleagues. They used the results from no fewer than 19 studies. They found no statistically significant link between intake of either lactose or individual dairy foods like milk, cheese and yogurt and the risk of ovarian cancer.
My conclusion is that on the balance of evidence that there is unlikely to be a link between consumption of yogurt, other dairy foods or galactose and the risk of developing ovarian cancer. If there is any causal link then it is so tiny an influence that it is at or below the limits of detection by the methods currently available to nutritional scientists. Given this conclusion, then there seems no possibility, either now or after many more similar studies, that any valid and unbiased recommendation could be made to women to alter their consumption of dairy foods in order to alter their risk of developing ovarian cancer. Ovarian cancer is a relatively uncommon condition whose frequency increases with age. Around 1 in 70 women will develop the condition during their lifetime and most cases occur after the menopause. For example in one large Dutch study with older (postmenopausal) women about 40 women per 10,000 developed the condition over the 11.3 years of follow-up. Current epidemiological methods cannot definitively detect small increases in this risk (say 20%) and confidently attribute it to an association with a dietary factor.
Despite the weakness of the evidence it contains, this paper by Daniel Cramer and his colleagues has helped to spawn hundreds of other papers over almost 3 decades without really advancing our understanding of the causes of ovarian cancer or our ability to make valid dietary recommendations to reduce it. Many more studies will probably be published in the future without any real prospect of changing that conclusion.
Sources used that do not have links in the text
Cramer, D.M. et al (1989) Galactose consumption and metabolism in relation to risk of ovarian cancer. Lancet 2, 66-71.
Liu, J. et al (2015) Milk, yogurt, and lactose intake and ovarian cancer risk: a meta-analysis. Nutrition and Cancer 67, 68-72. For those who have access to EBSCO: http://web.b.ebscohost.com/ehost/pdfviewer/pdfviewer?sid=f16891ef-c38c-4b37-adc4-1bebf2e4314b%40sessionmgr107&vid=1&hid=123
Stein, R (1989) Dairy Products linked to ovarian cancer. United Press International newswire July 12th 1989. Available through NEXIS: https://www.nexis.com/ .