A recent study in the journal Pediatrics reported that only 29% of clinical studies in children have been published. This finding reinforces previous studies that there is significant publication bias in paediatric studies. These findings are a cause for serious concern.

What is publication bias? Essentially, it is the selected publication of studies based on the results, such as only publishing studies that demonstrate a drug works while not publishing studies that demonstrate harms.

Publication bias is a serious problem in healthcare and can have a large influence on treatment decisions by only providing limited information. Researchers have demonstrated substantial publication bias in certain areas such as the antidepressant medication reboxetine.

Several initiatives have been spearheaded to help reduce publication bias. The creation of open-access journals have shifted the focus from the importance of the results (as judged by a journal editorial committee) to the methodological rigour by which the study was completed.

But more importantly has been the creation of online trial registries, such as ClinicalTrials.gov launched in 2000. These registries serve as central databases of all the current and on-going clinical studies. Registration is optional, however in 2005 the ICMJE made registration of clinical trials as a pre-requisite of publication. Although this does not represent all journals, it sent a strong message of the importance of registration.

However despite the creation of trial registries, less than half of US based National Institute of Health (i.e. government) funded trials in children were registered on ClinicalTrials.gov. Another important finding was the lack of information included on the registries. One-third of all clinical studies terminated early did not provide any information about why they were stopped. The situation was similar for suspended studies with one quarter not providing information.

Were these studies stopped because of harms? Were the investigators no longer able to recruit children to enroll? Whatever the reason the studies were stopped, this information must be made public.

Registration of all clinical studies involving children must be made mandatory. This is the only way to minimise publication bias and increase the reporting of research. This would create massive industry uproar, but is it ethical to enroll children in a clinical study without having it publicly registered? At a minimum any trial that receives government funded must be registered.

However registration of studies is only one element of the formula. What about the dissemination of the results? Less than 10% of completed studies in children had results posted and publicly available. With the low publication rates of registered studies, and the even lower rate of posting results, how much information is still missing?

Indeed progress has been made to increase the quality and transparency of clinical studies in children but more is needed. We cannot assume that because trial registries exist that they are being used. Complacency must be replaced with compliance. It seems that more often than not, the little ones have the biggest problems.

I have been concerned about the study that appeared in the BMJ recently about the association between industry affiliation and position on cardiovascular risk. Basically there was a clear and strong link between the orientations of authors’ expressed views on roiglitazone and their financial conflicts of interest: a drug that has concerned trusttheevidence previously.

In addition to this, an evaluation of solutions to sponsorship bias of more than 40 primary studies, and three recent systematic reviews and meta-analyses, have shown a clear association between pharmaceutical industry funding of clinical trials and pro-industry results.

What is alarming is that only half of the articles analysed in the BMJ had competing interest statements. However, this is much better that the 3% that was reported in 1998 by Stelfox. Authors at this time who supported the use of calcium-channel antagonists were significantly more likely than neutral or critical authors to have financial relationships with manufacturers of calcium-channel antagonists.

So in pondering this problem I think my solution is for primary studies, conflicts of interest should be disclosed when the article is published. However, for editorials and reviews where there is a direct conflict of interest (viz the researcher has received direct cash payments in the last five years) they should not be published by peer reviewed journals and should be left to those with an impartial viewpoint to publish in order to clear the muddy waters.

Understanding bias in clinical studies can help identify some of the reasons why we reach the wrong conclusions about the effects of interventions. Some of the biases we will be looking out for in 2010 include:

1. Publication bias: Positive findings are more likely to be published in medical journals than negative findings (Tamiflu). Media coverage of health issues often tends to be biased towards publication of stories which will grab headlines (swine flu).
2. Citation bias: This one is even more scary than publication bias: published articles tend to cite other articles that support their views rather than those articles which refute their views, and so have a negative impact on the scientific truth.
3. Selection bias: If the population studied is not representative of the population we want to draw conclusions about, then the study has a selection bias. The Framingham Heart study, for example, studied coronary heart disease in volunteers from a largely white, middle class population, and so we cannot necessarily draw conclusions about heart disease in blacks or other ethnic groups, or in populations in other countries.
4. Spectrum bias: A spectrum bias occurs when we overestimate how good a test is at picking up or excluding disease, because the test was evaluated in a biased sample of patients. The monofilament is a special tool, used to test whether diabetics have lost sensation in their feet. If it was only tested in people with mild sensory loss, then the monofilament may not be as good at picking up or excluding sensory loss, when it is used on people with severe sensory loss.
5. Information bias: If the way in which we measure an outcome or an exposure within a study is flawed, then we have an information bias. Data regarding carbon emissions rely on the integrity of countries and companies in their reporting. However, there is some evidence to suggest that such self-reporting (surprise surprise) is leading to gross under-reporting of emissions.
6. Recall bias: A special case of information bias is “recall bias”, where the ability of subjects to recall an exposure affects the results of a study. The difference between the amount of alcohol that people think that they drink and their actual alcohol consumption is a very good example.
7. Measurement bias: It is important to know whether the outcome measurement of interest is inaccurate. This can be due to inaccuracy in the measurement instrument or bias in the study participants expectations or responses. Often the way round the latter of these is to ensure adequate blinding.
8. Funding bias: an evaluation of solutions to sponsorship bias of more than 40 primary studies, and three recent systematic reviews and meta-analyses, have shown a clear association between pharmaceutical industry funding of clinical trials and pro-industry results. In 2010 elimination of such sponsorship bias should be a priority.

The final two biases are personal, in that when recognized, it may be possible to do something about them.

• Cognitive bias: is the tendency to make errors in judgment based on the way we think. In terms of diagnosis expertise is not a matter of acquiring an all-inclusive reasoning strategy, as several strategies may lead to the same diagnosis. These diagnoses are often correct; however, Clinicians tend to under-appreciate the likelihood their diagnoses are wrong and this tendency to overconfidence is related to both intrinsic and systemically reinforced factors.
• Reader Bias: Systematic errors of interpretation made during assumption by the user or reader of clinical information. These biases are due to the factors we put down to expertise: clinical experience, tradition, credentials, prejudice and human nature.

The last of these references by Richard Owen on reader bias is well worth a read, as it includes a whole host of further biases including: rivalry bias; personal habit bias, moral bias, clinical practice bias, do something bias. (The converse, do nothing bias, is common among academics), favoured design bias, prestigious journal bias, prominent author bias, famous institution bias (The converse: unrecognized institution bias), flashy title bias, friendship bias and my favourite “I am an epidemiologist" bias - Alternatively called bias bias - and is defined as repudiating a study containing any flaw in its design, analysis.

By Ami Banerjee and Carl Heneghan