We refer to Chapter 10 of the Cochrane Handbook.
The dissemination of research findings is not a division into published or unpublished, but a continuum ranging from the sharing of draft papers among colleagues, through presentations at meetings and published abstracts, to papers in journals that are indexed in the major bibliographic databases (Smith 1999). It has long been recognized that only a proportion of research projects ultimately reach publication in an indexed journal and thus become easily identifiable for systematic reviews.
Reporting biases arise when the dissemination of research findings is influenced by the nature and direction of results. Statistically significant, ‘positive’ results that indicate that an intervention works are more likely to be published, more likely to be published rapidly, more likely to be published in English, more likely to be published more than once, more likely to be published in high impact journals and, related to the last point, more likely to be cited by others. The contribution made to the totality of the evidence in systematic reviews by studies with non-significant results is as important as that from studies with statistically significant results.
The table below summarizes some different types of reporting biases.
Type of reporting bias
The publication or non-publication of research findings, depending on the nature and direction of the results
Time lag bias
The rapid or delayed publication of research findings, depending on the nature and direction of the results
The multiple or singular publication of research findings, depending on the nature and direction of the results
The publication of research findings in journals with different ease of access or levels of indexing in standard databases, depending on the nature and direction of results.
The citation or non-citation of research findings, depending on the nature and direction of the results
The publication of research findings in a particular language, depending on the nature and direction of the results
The selective reporting of some outcomes but not others, depending on the nature and direction of the results
While publication bias has long been recognized and much discussed, other factors can contribute to biased inclusion of studies in meta-analyses. Indeed, among published studies, the probability of identifying relevant studies for meta-analysis is also influenced by their results. These biases have received much less consideration than publication bias, but their consequences could be of equal importance.
In 1989, Gøtzsche found that, among 244 reports of trials comparing non-steroidal anti-inflammatory drugs in rheumatoid arthritis, 44 (18%) were redundant, multiple publications, which overlapped substantially with a previously published article. Twenty trials were published twice, ten trials three times and one trial four times (Gøtzsche 1989). The production of multiple publications from single studies can lead to bias in a number of ways (Huston 1996). Most importantly, studies with significant results are more likely to lead to multiple publications and presentations (Easterbrook 1991), which makes it more likely that they will be located and included in a meta-analysis. It is not always obvious that multiple publications come from a single study, and one set of study participants may be included in an analysis twice. The inclusion of duplicated data may therefore lead to overestimation of intervention effects, as was demonstrated for trials of the efficacy of ondansetron to prevent postoperative nausea and vomiting (Tramèr 1997).
Other authors have described the difficulties and frustration caused by redundancy and the ‘disaggregation’ of medical research when results from a multi-centre trial are presented in several publications (Huston 1996, Johansen 1999). Redundant publications often fail to cross-reference each other (Bailey 2002, Barden 2003) and there are examples where two articles reporting the same trial do not share a single common author (Gøtzsche 1989, Tramèr 1997). Thus, it may be difficult or impossible for review authors to determine whether two papers represent duplicate publications of one study or two separate studies without contacting the authors, which may result in biasing a meta-analysis of this data.
Research suggests that various factors related to the accessibility of study results are associated with effect sizes in trials. For example, in a series of trials in the field of complementary and alternative medicine, Pittler and colleagues examined the relationship between trial outcome, methodological quality and sample size with characteristics of the journals of publication of these trials (Pittler 2000). They found that trials published in low or non-impact factor journals were more likely to report significant results than those published in high-impact mainstream medical journals and that the quality of the trials was also associated with the journal of publication. Similarly, some studies suggest that trials published in English language journals are more likely to show strong significant effects than those published in non-English language journals (Egger 1997b), however this has not been shown consistently (Moher 2000, Jüni 2002, Pham 2005).
The term ‘location bias’ is also used to refer to the accessibility of studies based on variable indexing in electronic databases. Depending on the clinical question, choices regarding which databases to search may bias the effect estimate in a meta-analysis. For example, one study found that trials published in journals that were not indexed in MEDLINE might show a more beneficial effect than trials published in MEDLINE-indexed journals (Egger 2003). Another study of 61 meta-analyses found that, in general, trials published in journals indexed in EMBASE but not in MEDLINE reported smaller estimates of effect than those indexed in MEDLINE, but that the risk of bias may be minor, given the lower prevalence of the EMBASE unique trials (Sampson 2003). As above, these findings may vary substantially with the clinical topic being examined.
A final form of location bias is regional or developed country bias. Research supporting the evidence of this bias suggests that studies published in certain countries may be more likely than others to produce research showing significant effects of interventions. Vickers and colleagues demonstrated the potential existence of this bias (Vickers 1998).
The perusal of the reference lists of articles is widely used to identify additional articles that may be relevant although there is little evidence to support this methodology. The problem with this approach is that the act of citing previous work is far from objective and retrieving literature by scanning reference lists may thus produce a biased sample of studies. There are many possible motivations for citing an article. Brooks interviewed academic authors from various faculties at the University of Iowa and asked for the reasons for citing each reference in one of the authors’ recent articles (Brooks 1985). Persuasiveness, i.e. the desire to convince peers and substantiate their own point of view, emerged as the most important reason for citing articles. Brooks concluded that authors advocate their own opinions and use the literature to justify their point of view: “Authors can be pictured as intellectual partisans of their own opinions, scouring the literature for justification” (Brooks 1985).
In Gøtzsche’s analysis of trials of non-steroidal anti-inflammatory drugs in rheumatoid arthritis, trials demonstrating a superior effect of the new drug were more likely to be cited than trials with negative results (Gøtzsche 1987). Similar results were shown in an analysis of randomized trials of hepato-biliary diseases (Kjaergard 2002). Similarly, trials of cholesterol lowering to prevent coronary heart disease were cited almost six times more often if they were supportive of cholesterol lowering (Ravnskov 1992). Over-citation of unsupportive studies can also occur. Hutchison et al. examined reviews of the effectiveness of pneumococcal vaccines and found that unsupportive trials were more likely to be cited than trials showing that vaccines worked (Hutchison 1995).
Citation bias may affect the ‘secondary’ literature. For example, the ACP Journal Club aims to summarize original and review articles so that physicians can keep abreast of the latest evidence. However, Carter et al. found that trials with a positive outcome were more likely to be summarized, after controlling for other reasons for selection (Carter 2006). If positive studies are more likely to be cited, they may be more likely to be located and, thus, more likely to be included in a systematic review, thus biasing the findings of the review.
Reviews have often been exclusively based on studies published in English. For example, among 36 meta-analyses reported in leading English-language general medicine journals from 1991 to 1993, 26 (72%) had restricted their search to studies reported in English (Grégoire 1995). This trend may be changing, with a recent review of 300 systematic reviews finding approximately 16% of reviews limited to trials published in English; systematic reviews published in paper-based journals were more likely than Cochrane reviews to report limiting their search to trials published in English (Moher 2007). In addition, of reviews with a therapeutic focus, Cochrane reviews were more likely than non-Cochrane reviews to report having no language restrictions (62% vs. 26%) (Moher 2007).
Investigators working in a non-English speaking country will publish some of their work in local journals (Dickersin 1994). It is conceivable that authors are more likely to report in an international, English-language journal if results are positive whereas negative findings are published in a local journal. This was demonstrated for the German-language literature (Egger 1997b).
Bias could thus be introduced in reviews exclusively based on English-language reports (Grégoire 1995, Moher 1996). However, the research examining this issue is conflicting. In a study of 50 reviews that employed comprehensive literature searches and included both English and non-English-language trials, Jüni et al reported that non-English trials were more likely to produce significant results at P<0.05, while estimates of intervention effects were, on average, 16% (95% CI 3% to 26%) more beneficial in non-English-language trials than in English-language trials (Jüni 2002). Conversely, Moher and colleagues examined the effect of inclusion or exclusion of English-language trials in two studies of meta-analyses and found, overall, that the exclusion of trials reported in a language other than English did not significantly affect the results of the meta-analyses (Moher 2003). These results were similar when the analysis was limited to meta-analyses of trials of conventional medicines. When the analyses were conducted separately for meta-analyses of trials of complementary and alternative medicines, however, the effect size of meta-analyses was significantly decreased by excluding reports in languages other than English (Moher 2003).
The extent and effects of language bias may have diminished recently because of the shift towards publication of studies in English. In 2006, Galandi et al. reported a dramatic decline in the number of randomized trials published in German-language healthcare journals: with fewer than two randomized trials published per journal and year after 1999 (Galandi 2006). While the potential impact of studies published in languages other than English in a meta-analysis may be minimal, it is difficult to predict in which cases this exclusion may bias a systematic review. Review authors may want to search without language restrictions and decisions about including reports from languages other than English may need to be taken on a case-by-case basis.
In many studies, a range of outcome measures is recorded but not all are reported (Pocock 1987, Tannock 1996). The choice of outcomes that are reported can be influenced by the results, potentially making published results misleading. For example, two separate analyses (Mandel 1987, Cantekin 1991) of a double-blind placebo-controlled trial assessing the efficacy of amoxicillin in children with non-suppurative otitis media reached opposite conclusions mainly because different ‘weight’ was given to the various outcome measures that were assessed in the study. This disagreement was conducted in the public arena, since it was accompanied by accusations of impropriety against the team producing the findings favourable to amoxicillin. The leader of this team had received substantial fiscal support, both in research grants and as personal honoraria, from the manufacturers of amoxicillin (Rennie 1991). It is a good example of how reliance upon the data chosen to be presented by the investigators can lead to distortion (Anonymous 1991). Such ‘outcome reporting bias’ may be particularly important for adverse effects. Hemminki examined reports of clinical trials submitted by drug companies to licensing authorities in Finland and Sweden and found that unpublished trials gave information on adverse effects more often than published trials (Hemminki 1980). Since then several other studies have shown that the reporting of adverse events and safety outcomes in clinical trials is often inadequate and selective (Ioannidis 2001, Melander 2003, Heres 2006). A group from Canada, Denmark and the UK recently pioneered empirical research into the selective reporting of study outcomes (Chan 2004a, Chan 2004b, Chan 2005). These studies are described in Chapter 8 of the Handbook, along with a more detailed discussion of outcome reporting bias.
We highlight two recent publications
- The impact of outcome reporting bias in randomised controlled trials on a cohort of systematic reviews. Kirkham J.J. et.al. BMJ. 2010; 340:c365
- Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. Sterne, J.A.C. BMJ. 2011; 343:d4002