Today’s grossly overhyped health scare story comes courtesy of the BMJ’s Open Access journal and, in its original form, sports the following title: “Oral contraceptive use is associated with prostate cancer: an ecological study“.
Needless to say, by the time this paper filtered through into the mainstream press the coverage of its findings had taken on an altogether more alarmist tone:
Women using the contraceptive pill may have led to a global increase in prostate cancer, research shows.
Regions where the oral contraceptive was commonplace had higher incidents of the cancer, it was found.
Scientists say they cannot be sure what causes the link – but suggest that male exposure to oestrogen could heighten the risk.
Excess oestrogen exposure has been linked to cancer, and it is thought that widespread use of the pill may have led to more oestrogen, which is not easily broken down, entering the water supply and food chain.
That’s the Daily Telegraph. Over at the Daily Mail, where getting just the right tone on barely repressed hysteria is an essential feature of the paper’s business model, this story kicked off with the headline:
‘Contraceptive Pill linked to the rise in cases of Prostate Cancer’
…only for that to be junked within a matter of hours in favour of
‘Rise in prostate cancer ‘due to use of the Pill which increases men’s exposure to oestrogen’
And even the BBC has got in on the act with this entry in the ongoing series ‘Questions to which the answer is no!’:
Meanwhile, out in the wilds of the Twittersphere, Ed Yong offered up this pithy summation of the paper’s merits as a piece of serious research:
They say study is “hypothesis-generating”. Common refrain w/ ecol studies. It means “so shallow as to be useless”
Thus proving that you if you know where to look, you can easily find more sense in 140 characters than is evidence in the combined efforts of two national newspapers and (sadly) the BBC.
So why is Ed, who knows a thing or two about cancer research, so dismissive of this paper’s findings?
The answer, as you might expect, lies in the paper and the manner in which it has generated and then presented its findings. What the researchers have actually done here is take two datasets and carry out a couple of student-grade statistical comparisons in order to see if anything shakes loose by way of an association between the data, One of the datasets shows the estimated age-standardised prostate cancer incidence and mortality rates for 88 different countries, covering all five continents; the other gives an estimate of the usage of different types of contraception in those same countries including, of course, the contraceptive pill.
And that’s pretty much it but for an additional multivariable analysis in which GDP per capita is used as a crude proxy in an effort to control for differences in access to screening for prostate cancer between countries.
So what did they find?
In truth, not much – a statistically significant correlation between the incidence of prostate cancer and the prevalence of oral contraceptive use (r=0.61, p<0.05) which didn’t carry forward onto other forms of contraception or onto mortality rates and that’s pretty much it.
After issuing a stock warning that correlation does not necessarily mean causation:
This study represents the first systematic analysis of associations between OC use and prostate cancer. It is an ecological study and thus has, as with all correlational studies, significant limitations with respect to causal inference. As such, it must be considered hypothesis generating.
…the author’s moved swiftly onto their exercise in ‘hypothesis generating’*, i.e. a quick run through a few speculative ideas as to what might account for this apparent association between prostate cancer and oral contraceptives from which they go on to suggest, without any trace of irony, that more research is needed.
*Cynics amongst you might be inclined to suspect that the term ‘hypothesis generating’ is, here, likely to be synonymous with the term ‘research grant generating’ but I couldn’t possibly comment.
Now, ordinarily, I tend to be fairly indulgent of this kind of nonsense. It may well be the case that this study, for its obvious limitations, might help generate a few research dollars for a proper investigation of the author’s hypothetical link between environmental oestrogen and prostate cancer and even if any subsequent studies prove to be negative then we’ll at least have learned something and we can knock that hypothesis of the list and get on with investigating other possible causes.
However, on this occasion, any largesse I might otherwise be inclined to offer the authors is rather tempered by my awareness of the possible unintended consequences of the overhyped media coverage this paper is current receiving.
Quite a while ago, I knocked out a bit of quick and dirty research on the association between media coverage of sexual health issues and trends in teenage conception/abortion rates since the 1970, which I blogged over at Lib Con – sadly, the graph that I included in this article has gone AWOL, so you’ll have to trust that I’m accurately reporting my findings.
What I did at the time was relatively simple – I took a graph which plotted the long term trends in teenage conception, birth and abortion rates since the late 1960s/early 70s, when both abortion and the contraceptive pill became readily, and legally, available to women in the UK, and onto the graph I mapped any signficant period of heavy media coverage of sexual health related issues, i.e. abortion related legislation in parliament, periods of heighten anti-abortion activism and major health scare surrounding both STDs and the safety of oral and other contraceptives.
What emerged from this quite simple analysis was a very striking picture of media sensitivity in conception and abortion trends in young women. Put simply, when STDs – and particularly new STDs – are featured heavily in the media, then conception rates in young women start to fall shortly afterwards, and this has the knock on effect, of course, of generating a parallel fall in the abortion rate. The most striking example of this this phenomenon is, as you might well expect, the major HIV/AIDS campaign of the late 80s and early 90s but the same, if less marked trend, can also been seen in the data for the periods which followed a rash a media coverage of herpes (early 80s) and chlamydia (early 2000s)
Conversely, whenever there has been a major safety scare relating to oral contraception or IUDs (i.e. the Dalkon Shield), these have been followed shortly afterwards by a marked increase in teenage conceptions, a pattern which is consistent with young women switching the less reliable methods of contraception as a direct result of heavy media coverage of alleged safety issues/side effects of oral contraceptives.
Speculative hypotheses are all very well when they’re confined to the relatively safe environment of research literature, but when they leak out into the outside world there may be unforeseen consequences which researchers need to bear in mind when they’re writing up their research and, particularly, when they set about knocking press releases in an effort to generate publicity for their work.
This observation leads, naturally, to the question of whether this particular paper’s result stack up in scientific terms and my strong feeling is that they don’t.
To understand why, we need to reevaluate their data in terms of something that we know very well when it comes to prostate cancer and that something is simply that the risk of developing prostate cancer increases as men get older.
This is a potentially significant confounding factor for this study for the simple reason that the countries included in the dataset differ markedly not only in their incidence of prostate cancer and oral contraceptive use but also in their basic age demographics. This is controlled for in this study, to a degree, by the use of age standardised incidence and mortality rates which, at least in theory, should reduce the confounding influence of differences in the overall age profile of different countries.
In theory, if the researchers have controlled fully for age then a simple correlation between the age standardised incidence rates for a disease and standard age-based metric, such as life expectancy at birth, should produce an outcome somewhere close to a null hypothesis – this tells you both that you’ve eliminated age as a confounding factor and that, therefore, any observable effect in other correlation series you run against the same dataset is independent of any age-related effects.
So, bearing that in mind, I pulled down the same basic dataset as the authors and ran a couple of correlations against just about the most basic, and relevant, age-related metric I could think of; estimated male life expectancy at birth, from which I generated the following scatter diagrams:
The top graph, with the red trendline, is based on a full dataset which includes 176 countries, the bottom graph (green trendline) uses only the 88 countries included in this original paper and both graphs look extremely similar, i.e both show a trend in which the age standardised incidence of prostate cancer increases as male life expectancy increases, which is really telling us nothing that we don’t already know. The risk of developing prostate cancer increases with age and this inevitably results in a much lower incidence of prostate cancer in countries with low male life expectancy simply because a large proportion of men who might well have gone on to develop this particular cancer will dies of other causes long before it ever gets around to putting in an appearance.
What this also tells us is that age standardisation of the incidence statistic hasn’t done a particularly good job of controlling for the confounding effects of age. In terms of correlation coefficients for both graphs, the first graph with the full dataset has a coefficient (r) of 0.44 and a p-value <0.001 while the second graph has a coefficient of 0.49 and a similar p-value, ie. <0.001. Both are high side of moderate while the original studies correlation to oral contraceptive use is, at 0.61, on the low side of strong but – interestingly – just with the upper bounds of the 95% confidence interval for the correlation between incidence and life expectancy in the second graph, which is based on the same set of 88 countries.
Age remains a considerable source of confounding here and seems to be entirely capable of accounting for most, if not all of the effect reported in the original study. This doesn’t rule out the authors’ preferred environmental oestrogen hypothesis, of course, but it does suggest that it is, for the time being, a rather weak hypothesis unless, and until, it can be buttressed with solid evidence from other sources – and, to date, the evidence that is available seems to be somewhat equivocal as the paper, itself, notes:
Prostate cancer (PCa) is the most common male malignancy in the Western world, and risk factors associated with this cancer remain ill defined. The only acknowledged risk factors thus far are: age, ethnicity and family history. Several studies have suggested that oestrogen exposure may increase the risk of prostate cancer, while other studies have not found an association.
So what we appear to have here is not much more than a lot of handwaving on the back of very little evidence – the discussion section of the paper would just about pass muster as a crib sheet for an environmental science student looking to knock out a quick essay or term paper but otherwise it doesn’t seem add much to the overall sum of human knowledge.
To illustrate just exactly how little, I ran one final correlation on the full dataset of 176 countries for which I have age standardised incidence data for prostate cancer and after a little very minor tweaking to ensure that both sets of data I used ran in the correct order, I ran the calculation and got a result of 0.42, only just short the correlation score for incidence/male life expectancy.
So what did I test?
Well, put it this way – there appears to be a moderate correlation (and one not far short of the correlation with male life expectancy) between the incidence of prostate cancer and national sporting prowess as evidenced by international performance in the beautiful game.
Yes, that last test I ran used the current FIFA international rankings – ain’t science wonderful!