This is the third instalment of ‘The case for Open Research’ series of blogs exploring the problems with Scholarly Communication caused by having a single value point in research – publication in a high impact journal. The first post explored the mis-measurement of researchers and the second looked at issues with authorship.
This blog will explore the accuracy of the research record, including the ability (or otherwise) to reproduce research that has been published, what happens if research is retracted, and a concerning trend towards altering hypotheses in light of the data that is produced.
Science is thought to progress through the building of knowledge through questioning, testing and checking work. The idea of ‘standing on the shoulders of giants’ summarises this – we discover truth by building on previous discoveries. But scientists are very rarely rewarded for being right, they are rewarded for publishing in certain journals and for getting grants. This can result in distortion of the science.
How does this manifest? The Nine Circles of Scientific Hell describes questionable research practices that occur, ranging from Overselling, Post-Hoc storytelling, p-value Fishing, Creative use of Outliers to Non or Partial Publication of Data. We will explore some of these below. (Note this article appears in a special issue of Perspectives on Psychological Science on the Replicability in Psychological Science – which contains many other interesting articles).
Much as we like to think of science as an objective activity it is not. Scientists are supposed to be impartial observers, but in reality they need to get grants, and publish papers to get promoted to more ‘glamorous institutions’. This was the observation of Professor Marcus Munafo in his presentation ‘Scientific Ecosystems and Research Reproducibility’ at the Research Libraries UK conference held earlier this year (the link will take you to videos of the presentations). Monafo observed that scientists are rarely rewarded for being right, so the scientific record is being distorted by the scientific ecosystem.
Monafo, a Biological Psychologist at Bristol University, noted that research, particularly in the biomedical sciences, ‘might not be as robust as we might have hoped‘.
The reproducibility crisis
A recent survey of over 1500 scientists by Nature tried to answer the question “Is there a reproducibility crisis?” The answer is yes, but whether that matters appears to be debatable: “Although 52% of those surveyed agree that there is a significant ‘crisis’ of reproducibility, less than 31% think that failure to reproduce published results means that the result is probably wrong, and most say that they still trust the published literature.”
There are certainly plenty of examples of the inability to reproduce findings. Pharmaceutical research can be fraught. Some research into potential drug targets found that in almost two-thirds of the projects looked at, there were inconsistencies between published data and the data resulting from attempts to reproduce the findings.
There are implications for medical research as well. A study published last month looked at functional MRI (fMRI), noting that when analysing data using different experimental designs they should in theory find a significance threshold of 5% (a p-value of less than 0.05 which is conventionally described as statistically significant). However they found “the most common software packages for fMRI analysis (SPM, FSL, AFNI) can result in false-positive rates of up to 70%. These results question the validity of some 40,000 fMRI studies and may have a large impact on the interpretation of neuroimaging results.”
A 2013 survey of cancer researchers found that approximately half of respondents had experienced at least one episode of the inability to reproduce published data. Of those people who followed this up with the original authors, most were unable to determine why the work was not reproducible. Some of those original authors were (politely) described as ‘less than “collegial”’.
So what factors are at play here? Partly it is due to the personal investment in a particular field. A 2012 study of authors of significant medical studies concluded that: “Researchers are influenced by their own investment in the field, when interpreting a meta-analysis that includes their own study. Authors who published significant results are more likely to believe that a strong association exists compared with methodologists.”
This was also a factor in a study Why Most Published Research Findings Are False that considered the way research studies are constructed. This work found that “for most study designs and settings, it is more likely for a research claim to be false than true. Moreover, for many current scientific fields, claimed research findings may often be simply accurate measures of the prevailing bias.”
Psychology is a discipline where there is a strong emphasis on novelty, discovery and finding something that has a p-value of less than 0.05. There is such an issue with reproducibility in psychology that there are large efforts to try and reproduce psychological studies to estimate the reproducibility of the research. The Association for Psychological Science has launched a new article type of Registered Replication Reports which consists of “multi-lab, high-quality replications of important experiments in psychological science along with comments by the authors of the original studies”.
This is a good initiative, although there might be some resistance to this type of scrutiny. Something that was interesting from the Nature survey on reproducibility was the question of what happened when researchers attempted to publish a replication study. Note that only a few of respondents had done this, possibly because incentives to publish positive replications are low and journals can be reluctant to publish negative findings. The study found that “several respondents who had published a failed replication said that editors and reviewers demanded that they play down comparisons with the original study”.
What is causing this distortion of the research? It is the emphasis on publication of novel results in high impact journals. There is no reward for publishing null results or negative findings.
The p-value came up again in a discussion about HARKing at this year’s FORCE2016 conference (HARK stands for Hypothesising After the Results are Known – a term coined in 1998).
In his presentation at FORCE2016 Eric Turner, Associate Professor OHSU, spoke about HARKing (see this video 37 minutes onward). The process is that the researcher conceives the study, writes the protocol up for their eyes only, with a hypothesis and then collects lots of other data – ‘the more the merrier’ according to Turner. Then the researcher runs the study and analyses the data. If there is enough data, the researcher can try alternative methods and can play with statistics. ‘You can torture the data and it will confess to anything’ noted Turner. At some point the p-value will come out below 0.05. Only then does the research get written up.
Turner noted that he was talking about the kind of research where the work is trying to confirm a hypothesis (like clinical trials). This is different to hypothesis-generating research.
In the US clinical trials with human participants must be registered with the Federal Drug Agency (FDA) so it is possible to see the results of all trials. Turner talked about his 2008 study looking at antidepressant trials, where the journal version of the results supported the general view that antidepressants always beat placebo. However when they looked at the FDA version of all of the studies of the same drugs it happened that half of the studies were positive and half and half were not positive. The published record does not reflect the reality.
The majority of the negative studies were simply not published, but 11 of the papers had been ‘spun’ from negative to positive. These papers had a median impact factor of 5 and median citations of 68 – these were highly influential articles. As Turner noted ‘HARKing is deceptively easy’.
This perspective is supported by the finding that a researcher is more likely to falsely find evidence that an effect exists than to correctly find evidence that it does not. Indeed Munafo noted that over 90% of the psychological literature finds what it purports to set out to do. Either the research being undertaken is extraordinarily mundane, or something is wrong.
Increase in retractions
So what happens when it is discovered that something that is published is incorrect? Journals do have a system which allows for the retraction of papers, and this is a practice which has been increasing over the past few years. Research looking at why the number of retractions have increased found that it was partly due to lower barriers to publication of flawed articles. In addition papers are now being retracted for issues like plagiarism and retractions are now happening more quickly.
Retraction Watch is a service which tracks retractions ‘as a window into the scientific process’. It is enlightening reading with several stories published every day.
An analysis of correction rates in the chemical literature found that the correction rate averaged about 1.4 percent for the journals examined. While there were numerous types of corrections, chemical structures, omission of relevant references, and data errors were some of the most frequent types of published corrections. Corrections are not the same as retractions, but they are significant.
There is some evidence to show that the higher the impact factor of the journal a work is published in, the higher the chance it will be retracted. A 2011 study showed a direct correlation between impact factor and the number of retractions, with New England Journal of Medicine topping the list. This situation has led to claims that the top ranking journals publish the least reliable science.
A study conducted earlier this year demonstrated that there are no commonly agreed definitions of academic integrity and malpractice. (I should note that amongst other findings the study found 17.9% (± 6.1%) of respondents reported having fabricated research data. This is almost 1 in 5 researchers. However there have been some strong criticisms of the methodology.)
There are questions about how retractions should be managed. In the print era it was not unheard of for library staff to put stickers into printed journals notifying a retraction. But in the ‘electronic age’ asked one author in 2002 when the record can be erased, is this the right thing to do because erasing the article entirely is amending history. The Committee on Publication Ethics (COPE) do have some guidelines for managing retractions which suggest the retraction be linked to the retracted article wherever possible.
However, from a reader’s perspective, even if an article is retracted this might not be obvious. In 2003* a survey of 43 online journals found 17 had no links between the original articles and later corrections. When present, hyperlinks between articles and errata showed patterns in presentation style, but lacked consistency. There are some good examples – such as Science Citation Index but there was a lack of indexing in INSPEC, and a lack of retrieval with SciFinder Scholar.
[*Note this originally said 2013, amended 2 September 2016]
All of this paints a pretty bleak picture. In some disciplines the pressure to publish novel results in high impact journals results in the academic record being ‘selectively curated’ at best. At worst it results in deliberate manipulation of results. And if mistakes are picked up there is no guarantee that this will be made obvious to the reader.
This all stems from the need to publish novel results in high impact journals for career progression. And when those high impact journals can be shown to be publishing a significant amount of subsequently debunked work, then the value of them as a goal for publication comes into serious question.
The next instalment in this series will look at gatekeeping in research – peer review.
Published 14 July 2016
Written by Dr Danny Kingsley