Falsification of research data
Falsification of research results
The ALLEA Code defines falsification as: manipulating research materials, equipment, images, or processes, or changing, omitting, or suppressing data or results without justification.
Falsification or fabrication?
There is often confusion between falsification and fabrication. In general, if results report on experiments that never took place, it is fabrication (making up results). If the experiments did take place, but the results were altered, it is falsification. However, there are some borderline cases: if the same image is used to represent two different experiments, this could be falsification (changing the data), or fabrication (making the data up). The difference may be significant for investigators and other scientists: if the ‘real’ data from a falsified experiment is available somewhere, it may be possible to correct the scientific record based on this data. If the data has been entirely fabricated, it is essentially worthless. However, this does not make falsification any less serious than fabrication.
Impact
The impact of falsification on science is similar to the impact of fabrication: it causes damage to the scientific record and to the careers of researchers involved. It will also damage the careers of collaborating researchers who were not aware of the falsification. Falsification imposes costs in terms of wasted research resources and the cost of investigating cases.
In contrast to fabrication, falsification also has negative impacts on research participants (since falsified results are based on real experiments). For human participants, this may range from wasted time and damage to their faith in science, to undergoing distressing or painful procedures for no good reason. This was particularly notable in the Wakefield case, where nurses were required to carry out painful procedures on children. Laboratory animals will also suffer for no good reason if they are used for experiments for which the results are later falsified.
Image manipulation
Although it is sometimes justifiable to alter images as discussed in Module 3, some types of image manipulation are fraudulent.
In the video below, Dr. Thorsten Beck of the Humboldt-Elsevier Advanced Data and Text Centre (HEADT Centre) in Berlin talks about some of the fundamental challenges of dealing with image manipulations.
Detection of potential mistakes, fabrication and/or fabrication
Similar to the solutions used by many journals and institutions to detect plagiarism, there is a variety of commercial and non-commercial tools and strategies available to help detecting other specific irregularities. Some examples:
- Tools to detect potential issues with the statistical analyses in studies or the underlying dataset:
- The Grim test evaluates whether the reported means of summary results are consistent with the given sample size and number of items (Brown and Heathers, 2017). Simply put, when working with whole numbers and a fixed number of observations, some means cannot be obtained, indicating there is something wrong with the analysis and/or the underlying dataset.
- Statcheck checks whether the reported p-values are consistent with the accompanying test statistic and degrees of freedom and flags results where the computed p-values don’t match as an error (Nuijten et al, 2016).
- The tortured phrases detector may be used to detect unexpected weird phrases which could be the results of questionable AI-generated or rewritten texts attempting to evade plagiarism detection software (Cabanac, 2021).
- Within the biomedical field, Seek & Blastn may help to verify the nucleotide sequence reagents in publications and manuscripts (Labbé, 2019).
- There are multiple commercial tools available to assist in the detection of alterations and duplications of scientific images. An overview of interesting tools has been compiled by the STM Image Alterations & Duplications Working Group.
Although the above tools have great value in detecting potential issues, it is, however, important to note that none of these tools are 100% accurate and may therefore create both false positive as false negative hits. Validity of the observations always has to be checked in person. Moreover, a confirmed hit does not automatically mean falsification or fabrication have occurred. Mistakes because of honest errors cannot be ruled out.