Definition of bias
The word "bias" is derived from the French biais, meaning "askew" (Store medisinske leksikon, Sverre Braut). "Bias" has no unambiguous or generally accepted meaning, either inside or outside the framework of research ethics. To render it meaningful, "bias" must be linked to another word (such as selection bias, recall bias and publication bias). In Norwegian, the word "skjevhet" (skewness) is often used instead of the English term "bias", as recommended in the advice to authors of the Journal of the Norwegian Medical Association (www.tidsskriftet.no). Likewise in Norwegian, "skjevhet" must be used in a compound term to give any meaning (such as selection "skjevhet", recall "skjevhet" and publication "skjevhet"). In this article, the English and Norwegian terms will be used interchangeably.
The Norwegian Large Encyclopaedia of Medicine refers to the concept of bias as: "used in statistical and empirical research when results or inferences systematically deviate from the true values. Bias may occur as a result of errors or inaccuracies in the sample of study subjects, choice of method of investigation or assessment of results". The American National Library of Medicine defines bias as: "any factor that distorts the true nature of an event or observation. In clinical investigations, a bias is any systematic factor other than the intervention of interest that affects the magnitude of (i.e. tends to increase or decrease) an observed difference in the outcomes of a treatment group and a control group. Bias diminishes the accuracy (though not necessarily the precision) of an observation. Randomisation is a technique used to decrease this form of bias. Bias also refers to a prejudiced or partial viewpoint that would affect someone's interpretation of a problem. Double blinding is a technique used to reduce this type of bias." (www.nlm.nih.gov).
Bias can also be defined simply as "a form of systematic error that may have an effect on scientific studies and distort the measurement processes" (Sica GT, 2006)
Types of bias
In research there are many types of bias, for which there is no generally accepted categorisation. Many forms may overlap. Bias may occur at all stages of a research project, such as during generation of hypotheses, planning, implementation and funding of studies, during collection, processing and interpretation of study data, as well as during publication (dissemination) of research data. This section will describe some common forms of such systematic error in research studies.
This may include selection bias, sample bias, loss-to-follow-up bias, disease spectrum bias, referral bias, participation (response) bias, self-selection bias (i.e. that volunteer participants in a research study will often differ from those who do not accept the offer to participate). The latter will also apply to simple records-based retrospective studies (observational studies) that are based on the patients' consent to access to their medical records (Kho ME et al. 2009), a common research ethics requirement for use of confidential information collected by the health services. Kho et al. argue that researchers should have the opportunity to "request a waiver of consent from the research ethics board" (Kho ME et al, 2009) wherever such a requirement for consent may result in biased study results. Such a waiver presumes that the retrieval and use of the data do not represent a threat to the privacy of individual patients or the patient group.
The choice of study population is essential to be able to generalise the research results. The selection of respondents/study participants may be affected by a number of factors. The final participation in the study depends on a series of factors that individually or jointly may give rise to bias in the interpretation of results with regard to the population that the study is intended to represent. Properties of the participants (and the phenomenon under investigation), and the interviewers may also give rise to bias, for example in the form of:
- recall bias (i.e. that the participants may have a varying ability to recall the phenomenon under investigation)
- interviewer bias
- follow-up bias
- response bias
- attrition bias etc.
Methodological bias also includes:
- scale bias, meaning that different people may have different norms regarding what the scale of a phenomenon is in fact measuring, such as the degree of happiness, pain etc.
The phenomenon of selection bias encompasses more than skewness in the selection of the study population (Odierna DH et al. 2013), such as the potential for unbalanced sampling of biological material, for example from a cancerous neoplasm (Ransohoff DF and Gourlay ML, 2010).
To enable the reader of a scientific publication to assess whether selection bias is present, it is essential that the researchers provide thorough information on inclusion and exclusion criteria, as well as the population background.
Epidemiological bias / statistical bias
Confounding is a common source of error in medical and health research (see e.g. Lang TA and Secic M, 2006). In epidemiological language, bias may result in incorrect estimates of associations, i.e. the observed study results may tend to be wrong and differ from the "true" results. One difference between random errors and systematic errors (such as selection bias) is that increasing the study population will reduce random errors down towards zero, while systematic errors cannot be reduced by increasing the sample size (Rothman KJ, 2002).
Publication/ dissemination bias
Imbalance in what research studies are published is often associated with the tendency of the pharmaceutical industry to restrict their publication of results to studies that benefit their own company. However, publication bias (skewed dissemination) may affect all types of research projects, not only pharmaceutical research, and occurs in qualitative and quantitative research alike.
The concept of publication bias also includes the phenomena:
- language bias (e.g. selective inclusion of studies published in English)
- access bias/cost bias (selective inclusion of studies that are easily/freely available or available at a low cost)
- familiarity bias (selective inclusion of studies from one's own specialty)
- outcome bias (selective reporting of individual results from a primary study but excluding other end points, where publication is often guided by the direction of the results and their statistical significance)
The choice of research studies to be included in a meta-analysis is an example of potential for publication bias, and this is illustrated in the case presented at the end of this article. (Direct link). A study found that 50% of all available drug studies may be absent from meta-analysis studies (Jørgensen AW et al. 2006). An example of the bias this could cause is presented in a study showing that meta-analyses of anti-hypertension drugs were positive in 91% of the studies that had ties to the pharmaceutical industry, compared to 72% of studies that had no such ties (Jørgensen AW et al. 2006). Meta-analyses that were supported by the pharmaceutical industry have also been found to be less transparent and to contain fewer reservations concerning the methodological limitations of the studies, as well as having more favourable conclusions than equivalent Cochrane analyses (not supported by the pharmaceutical industry) (Jørgensen AW et al. 2006).
A Cochrane review concluded that drug studies funded by the pharmaceutical industry overestimated effectiveness and underestimated possible adverse effects (Lundh et al. 2012). Sources of research funding and possible conflicts of interest may thus be seen as representing a possible risk of bias.
Publication bias also includes analysis and reporting bias (Dwan K et al. 2014), and in 40–62% of the publications in the field of health interventions the "primary outcome" had been changed in the publication of the study from the original protocol text (Dwan K et al. 2013). It is alarming to note that a study that re-analysed RCT (randomised clinical trial) data found a conclusion other than the one in the original publication in as much as 35% of the studies (Ebrahim S et al. 2014).
Numerous investigations have shown that positive research studies will have a greater likelihood of being published than negative ones, and often also more rapidly than studies that do not present such "positive" results (Hopewell S et al., 2009; Smith R, 2006). An article that describes beneficial effects of a new drug will have a greater likelihood of being published, within a shorter time, than one referring to a drug that has been tested without any positive effects on the recipient group. Reputable journals are concerned with publishing "negative" as well as "positive" research results, as emphasised by a former editor of the British Medical Journal: "If the question is important and the answer valid, then it mustn't matter whether the answer is positive or negative" (Smith R, 2006). He also believes that the researchers themselves help bring about such skewed publication of "positive" research results, by frequently not submitting "negative" studies for publication at all. One way to reduce such a publication bias of negative and positive studies could be to make more widespread provisions for pre-publication of the study protocol.
Today, most reputable journals require registration of intervention studies (such as drug studies) in an international database (such as the clinical trials system). A corresponding requirement for limited publication of the study results in such open databases after completion provides a more transparent access to the study results and may help reduce the problem of publication bias (Allison DB, 2009). This provides a better opportunity to assess the scope of those studies and results that fail to make their way into generally available research publications. Since 2008, this mandatory publication of study results applies to researchers who are under American jurisdiction (Bretthauer M and Haug C, 2009).
Reviewer bias can be regarded as part of the problems associated with publication (dissemination) bias. Biases and opinions held by a peer reviewer may affect the academic assessment of the research study, and thus the opportunity to have it published in a widely read journal and to affect the implementation of any findings made.
Consequences of bias
As a consequence of bias during the research process, the results obtained may not reflect reality, because of a "skewed" methodology and/or presentation of the results. Bias may to a varying degree reduce the relevance and applicability of a research study, since the study's conclusions are untrustworthy. At worst, such biased studies may give rise to recommendations for interventions, examinations and treatments that harm the targeted population or individuals, because these recommendations rest on a false or skewed premise. In health research, for example, bias may thus give rise to wrong priorities and health recommendations, because the study's conclusions have been drawn on an erroneous basis. This may have negative health consequences and be a waste of society's resources, with negative implications for individuals. Bias may also cause harm to future research projects, since future studies may come to be based on false assumptions. Moreover, when detected, bias may harm the public esteem of the research community and society's willingness to fund research projects.
The epidemiologist J.P. Ioannidis has published a number of studies demonstrating that bias is widespread and complex in biomedical research (Ioannidis JP, 2008) and that this may give rise to erroneous conclusions with potentially significant consequences for individuals and society. He summarises: "with increasing bias, the chance that the research findings are true diminishes considerably" (Ioannidis JP 2005).
How to detect bias in a research article
The opportunities for readers to check for bias in research studies are described in STARD (the STAndards for Reporting of Diagnostic Accuracy, Bossuyt PM et al., 2003). Other authors have provided suggestions for how authors' personal bias (conflicts of interest) can be checked for and detected after the fact in studies that can be assumed to have ties to the pharmaceutical industry (Gogol M, 2006).
The Cochrane Review Groups have also published recommendations for how readers may assess possible bias in systematic review studies (Lundh A and Gøtzsche PC, 2008). Proposals have also been made for models that use mathematical/statistical methods to detect publication bias resulting from a skewed selection of studies and study results in meta-analyses, for example the use of "funnel plots" that may provide a graphic indication of possible inclusion of a biased sample of studies (Rothstein HR et al., 2005).
How can bias be avoided/reduced? Raising awareness among researchers, journals and in society
The possibilities for bias in a study and how these can be avoided need to be considered already at the planning stage of a research project. Bias can be a result of deliberate as well as unintentional acts and choices, and in the former case it may also border on deliberate falsification. Avoiding all forms of bias in a research project, such as selection bias, is likely to be impossible. Other forms of bias may be easier to prevent as well as more essential in terms of research ethics, such as systematic exclusion of studies that do not corroborate the researcher's own hypotheses. Bias in the form of confounding may to some extent be corrected for in the statistical analyses, while certain other forms of bias do not lend themselves to adjustment, and may hence invalidate the study results.
Some forms of bias cannot be easily influenced directly by the researcher, such as the propensity among journals to accept articles with positive findings. As a researcher, one may nevertheless be able to help reduce this publication bias by acknowledging high-quality studies for their design when called upon to act as a peer reviewer of an article, and resisting the temptation to give recognition only to studies that demonstrate differences between therapies/ interventions/groups.
So-called personal "conflicts of interest" (COI) will often be assessed differently by different parties/people and comprise a lot more than merely financial ties, for example to a pharmaceutical enterprise. Friendships, business connections, family ties, political attitudes and personal opinions may all entail possibilities for bias in research methodologies. In general, conflicts of interest should be clearly declared in all publications to permit the readers to form an opinion as to whether authors may be considered to be biased in their approach or publication from a research area. The Vancouver group has prepared a new COI form which is more comprehensive than previous forms used for declaration of conflicts of interest (Drazen JM et al., 2010). Most reputable journals are expected to use such expanded self-reporting of possible conflicts of interest, thus to increase awareness and transparency related to personal bias.
All researchers ought to be aware of the phenomenon of bias in their own research practice as well as in their assessments of research conducted by others, since bias may give rise to erroneous conclusions. As a phenomenon, bias is found not only in the research world, but to a great extent also in society in general, in politics and the media. In general, researchers will to a greater or lesser extent be prone their own biases. Most researchers are deeply entrenched in a research tradition and social context of which they may not be fully aware. This may entail a comparatively significant risk of falling prey to methodological prejudices. The researcher's personal attitudes and awareness of the possibilities for prejudices and research biases are a key precondition for reducing the risk of bias, as well as for reducing any negative effects resulting from research bias. Bias can hardly be completely avoided in any research project. However, openness (transparency) in the presentation of research methodologies as well as any personal ties is essential for permitting society and other researchers to assess the extent and any possible effects of bias in a research project.
This article has been translated from Norwegian by Erik Hansen, Akasie språktjenester AS.