Systematic Review of the Empirical Evidence of Study Publication Bias and Outcome Reporting Bias

Background The increased use of meta-analysis in systematic reviews of healthcare interventions has highlighted several types of bias that can arise during the completion of a randomised controlled trial. Study publication bias has been recognised as a potential threat to the validity of meta-analysis and can make the readily available evidence unreliable for decision making. Until recently, outcome reporting bias has received less attention. Methodology/Principal Findings We review and summarise the evidence from a series of cohort studies that have assessed study publication bias and outcome reporting bias in randomised controlled trials. Sixteen studies were eligible of which only two followed the cohort all the way through from protocol approval to information regarding publication of outcomes. Eleven of the studies investigated study publication bias and five investigated outcome reporting bias. Three studies have found that statistically significant outcomes had a higher odds of being fully reported compared to non-significant outcomes (range of odds ratios: 2.2 to 4.7). In comparing trial publications to protocols, we found that 40–62% of studies had at least one primary outcome that was changed, introduced, or omitted. We decided not to undertake meta-analysis due to the differences between studies. Conclusions Recent work provides direct empirical evidence for the existence of study publication bias and outcome reporting bias. There is strong evidence of an association between significant results and publication; studies that report positive or significant results are more likely to be published and outcomes that are statistically significant have higher odds of being fully reported. Publications have been found to be inconsistent with their protocols. Researchers need to be aware of the problems of both types of bias and efforts should be concentrated on improving the reporting of trials.


Introduction
Study publication bias arises when studies are published or not depending on their results; it has received much attention [1,2]. Empirical research consistently suggests that published work is more likely to be positive or statistically significant (P,0.05) than unpublished research [3]. Study publication bias will lead to overestimation of treatment effects; it has been recognised as a threat to the validity of meta-analysis and can make the readily available evidence unreliable for decision making. There is additional evidence that research without statistically significant results takes longer to achieve publication than research with significant results, further biasing evidence over time [4][5][6]29]. This ''time lag bias'' (or ''pipeline bias'') will tend to add to the bias since results from early available evidence tend to be inflated and exaggerated [7,8].
Within-study selective reporting bias relates to studies that have been published. It has been defined as the selection on the basis of the results of a subset of the original variables recorded for inclusion in a publication [9]. Several different types of selective reporting within a study may occur. For example, selective reporting of analyses may include intention-to-treat analyses versus per-protocol analyses, endpoint score versus change from baseline, different time points or subgroups [10]. Here we focus on the selective reporting of outcomes from those that were originally measured within a study; outcome reporting bias (ORB).
Randomised controlled trials (RCTs) are planned experiments, involving the random assignment of participants to interventions, and are seen as the gold standard of study designs to evaluate the effectiveness of a treatment in medical research in humans [11]. The likely bias from selective outcome reporting is to overestimate the effect of the experimental treatment.
Researchers have considered selective outcome reporting to be a major problem, and deserving of substantially more attention than it currently receives [12]. Recent work [13][14][15][16][17][18][19] has provided direct empirical evidence for the existence of outcome reporting bias. Studies have found that statistically significant results had a higher odds of being fully reported compared to non-significant results for both efficacy and harm outcomes. Studies comparing trial publications to protocols are accumulating evidence on the proportion of studies in which at least one primary outcome was changed, introduced, or omitted.
Thus, the bias from missing outcome data that may affect a meta-analysis is on two levels: non-publication due to lack of submission or rejection of study reports (a study level problem) and the selective non-reporting of outcomes within published studies on the basis of the results (an outcome level problem). While much effort has been invested in trying to identify the former [2], it is equally important to understand the nature and frequency of missing data from the latter level.
The aim of this study was to review and summarise the evidence from empirical cohort studies that have assessed study publication bias and/or outcome reporting bias in RCTs approved by a specific ethics committee or other inception cohorts of RCTs.

Study inclusion criteria
We included research that assessed an inception cohort of RCTs for study publication bias and/or outcome reporting bias. We focussed on inception cohorts with study protocols being registered before start of the study as this type of prospective design were deemed more reliable. We excluded cohorts based on prevalence archives, in which a protocol is registered after a study is launched or completed, since such cohorts can already be affected by publication and selection bias.
Both cohorts containing exclusively RCTs or containing a mix of RCTs and non-RCTs were eligible. For those studies where it was not possible to identify the study type (i.e. whether any included studies were RCTs), we attempted to contact the authors to try to resolve this. In cases where it could not be resolved, studies were excluded. Those studies containing exclusively non-RCTs were excluded.
The assessment of RCTs in the included studies had to involve comparison of the protocol against all publications (for outcome reporting bias) or information from trialists (for study publication bias).

Search strategy
The first author (KD) alone conducted the search. No masking was used during the screening of abstracts. MEDLINE (1950MEDLINE ( to 2007, SCOPUS (1960 to 2007) and the Cochrane Methodology Register (1898 to 2007) were searched without language restrictions (final search December 2007 -see Appendix S1 for all search strategies). SCOPUS is a much larger database than EMBASE, it offers more coverage of scientific, technical, medical and social science literature than any other database. Over 90% of the sources indexed by EMBASE are also indexed by SCOPUS plus many other indexed sources as well.
Additional steps were taken to complement electronic database searches: First, the references given in the empirical evidence section of the HTA report of Song et al [1] were checked for relevance. Second, the lead reviewer of the protocol on the Cochrane library entitled 'Publication bias in clinical trials' [20] (Sally Hopewell) was contacted in November 2007 for references to studies included and excluded in their review. Their search strategy was compared to our own and differences in included studies were discussed between PRW, KD and Sally Hopewell. Finally, the lead or contact authors of all identified studies were asked to identify further studies.

Quality assessment
To assess the methodological quality of the included studies, we applied the same criteria as a recent Cochrane review [20]. In addition, we examined whether protocols were compared to publications in those studies that purported to investigate outcome reporting bias.

Data extraction
A flow diagram ( Figure 1, text S1) to show the status of approved protocols was completed for each empirical study by the first author (KD) using information available in the publication or further publications. Lead or contact authors of the empirical studies were then contacted by email and sent the flow diagram for their study to check the extracted data along with requests for further information or clarification of definitions if required. No masking was used and disagreements were resolved through discussion between KD and the lead or contact author of the empirical studies. Where comments from the original author were not available, PRW reviewed the report and discussed queries with KD.
Characteristics of the cohorts were extracted by the first author for each empirical study and issues relating to the methodological quality of the study were noted. We recorded the definitions of 'published' employed in each empirical study. Further, we looked at the way the significance of the results of the studies in each cohort were investigated (i.e. direction of results and whether the study considered a p-value #0.05 as definition of significance and where there were no statistical tests whether the results were categorised as negative, positive, important or unimportant). We extracted data on the number of positive and negative trials that were published in each cohort and we extracted all information on the main objectives of each empirical study and separated these according to whether they related to study level or outcome level bias.

Data analysis
This review provides a descriptive summary of the included empirical studies. We refrained from statistically combining results from the different cohorts due to the differences in their design.

Search results
The search of MEDLINE, SCOPUS and the Cochrane Methodology Register led to 973, 1717 and 554 references, respectively. Titles were checked by the first author (KD) and abstracts obtained for 57 potentially relevant studies. Abstracts were assessed for eligibility by the first author; 38 were excluded and full papers were obtained for 16. Only meeting abstracts were available for three studies [17,18,21] and their authors were contacted. Copies of their presentations were received and relevant data extracted.
The HTA report of Song et al [1] led to four potentially eligible empirical studies [3,4,26,27], all of which had been identified previously. References from the included empirical studies led to another paper [33] which gave extra information on the type of publication (full, abstract, none or unknown) for four eligible empirical studies [3,4,26,27]. The reference list provided by Sally Hopewell did not lead to any further studies.
Through contact with the authors, one reference [30] was located and found to be eligible and another [34] was identified  , About half of drug trials are not published. A high prevalence of prespecified outcomes are not reported and discrepancies includes primary outcomes. Completeness of reporting of an outcome is associated with statistical significance. 1 Easterbrook et al assumed that only studies that had been analysed had the potential for being written up and published, so tests for study publication bias were restricted to these. 2 Studies for which there was a full interview by the researchers of the cohort study and for which information on the nature of results and publication was provided. 3 Of the 520 studies with completed questionnaires, 321 had analysis undertaken with results available and were included in further analysis of the association between study outcome and time to publication. 4 Both groups are not mutually exclusive. 4% had a statement of both sources of funding. In the remainder, the protocols did not include information on how study was funded.
NI that gave more information on one of the eligible studies [5]. Thus in total, the search strategy identified 16 eligible empirical studies ( Figure 2). We are aware of three further empirical studies currently underway in Italy (D'Amico, personal communication), Germany (Von Elm, personal communication) and the USA (Djulbegovic, personal communication), but no further information is available at this stage.
However, two of these empirical studies [28,31] did not consider whether a study was submitted for publication.
Four cohorts included only RCTs [3,5,21,28]; in the remaining seven cohorts [4,26,27,[29][30][31][32] the proportion of included RCTs ranged from 14% to 56%. The results presented in the flow diagrams relate to all studies within each cohort because it was not possible to separate information for different types of studies (RCTs versus other).
Outcome reporting bias. Five empirical studies covered the entire process from the study protocol to the publication of study outcomes [13][14][15]17,18]. However, three of these empirical studies  [15]. doi:10.1371/journal.pone.0003081.g003 [13,17,18] did not consider whether a study was submitted for publication. Four cohorts included only RCTs [14,15,17,18]; in the remaining cohort [13] the proportion of included RCTs was 13%. Table 1 contains information on empirical study characteristics. The majority of the empirical study objectives related to study publication bias or outcome reporting bias.

Study Characteristics
Study publication bias. Three of the empirical studies investigating study publication bias also assessed time lag bias [4,5,29], one [28] assessed the outcome of protocols submitted to a research ethics committee (for example whether trials were started and if they were published) and another considered whether absence of acknowledged funding hampered implementation or publication [30]. Seven of the empirical studies [4,[26][27][28][29][30]32] assessed protocols approved by ethics committees, one [3] assessed those approved by health institutes, one assessed trials processed through a hospital pharmacy [21], one assessed studies funded by the NHS and commissioned by the North Thames Regional Office [31] and one empirical study [5] assessed trials conducted by NIH-funded clinical trials groups. The time period between protocol approval and assessment of publication status varied widely (less than one year to 34 years). Outcome reporting bias. Four of the empirical studies [13,15,17,18] assessed protocols approved by ethics committees and one empirical study [14] assessed those approved by a health institute. The time period between protocol approval and assessment of publication status varied from four to eight years.

Quality Assessment
Details of the methodological quality are presented in Table 2. The overall methodological quality of included empirical studies was good, with more than half of studies meeting all criteria. Study publication bias. Four of the eleven empirical studies [5,21,27,28] met all four of the criteria for studies investigating study publication bias (inception cohort, complete follow up of all trials, publication ascertained through personal contact with the investigator and definition of positive and negative findings clearly defined). In five empirical studies [3,4,26,29,30] there was less than 90% follow up of trials and in 2 empirical studies [31,32] the definition of positive and negative findings was unclear.
Outcome reporting bias. All five empirical studies [13][14][15]17,18] met all five criteria for studies investigating ORB (inception cohort, complete follow up of all trials, publication ascertained through personal contact with the investigator, definition of positive and negative findings clearly defined and comparison of protocol to publication).
As some studies may have several specified primary outcomes and others none, we looked at how each of the empirical studies dealt with this: Hahn et al [13] looked at the consistency between protocols and published reports in regard to the primary outcome and it was only stated that there were 2 primary outcomes in one study. In both of their empirical studies Chan et al [14,15] distinguished harm and efficacy outcomes but did consider the consistency of primary outcomes between protocols and publica- tions and stated how many had more than one primary outcome. Ghersi et al [17] included studies with more than one primary outcome and included all primary outcomes in the analysis but excluded studies with primary outcomes that were non identifiable or included more than 2 time points. This is due to complex outcomes being more prone to selective reporting. von Elm et al [18] considered harm and efficacy outcomes and primary outcomes.

Flow diagrams
The flow diagrams (Figures 3 to 18) show the status of approved protocols in included empirical studies based on available publications and additional information obtained such as number of studies stopped early or never started.
Study publication bias. No information other than the study report was available for one empirical study [26] due to its age. Information could not be located for three empirical studies [3,27,32]. A conference abstract and poster was only available for one empirical study presented over 10 years ago [21]. Extra information from lead or contact authors was available for six empirical studies [4,5,[28][29][30][31], including data to complete flow diagrams, information on definitions and clarifications.
Outcome reporting bias. A conference presentation only was available for one empirical study which is still to be published in full [17]. Extra information from lead or contact authors was available for four empirical studies [13][14][15]18], including data to complete flow diagrams, information on definitions, clarifications and extra information on outcomes. Original flow diagrams and questions asked are available on request. Figure 3 shows for illustrative purposes the completed flow diagram for the empirical study conducted by Chan et al [15] on the status of 304 protocols approved by the Scientific-Ethical Committees for Copenhagen and Frederiksberg in 1994-1995. The empirical study was conducted in 2003, which allowed sufficient time for trial completion and publication. Thirty studies were excluded as the files were not found. Surveys were sent to trial investigators with a response rate of 151 out of 274 (55%); of these two were ongoing, 38 had stopped early, 24 studies had never started and 87 studies were completed. Information from the survey responses (151) and the literature search alone (123) indicated that 120 studies had been submitted for publication and 154 studies had not been submitted for publication. Of the 120 submitted studies; 102 had been fully published, 16 had been submitted or were under preparation and two had not been accepted for publication. This resulted in 156 studies not being published.

Publication and trial findings
Study publication bias. Table 3 shows the total number of studies published in each cohort which varies widely from 21% to 93%. Nine of the cohorts [3][4][5]21,26,27,29,30,32] consider what proportion of trials with positive and negative results are published, ranging from 60% to 98% and from 19% to 85%, respectively. Only four cohorts [4,26,29,32] consider what percentage of studies with null results (no difference observed between the two study groups, p.0.10, inconclusive) are published (32% to 44%). The results consistently show that positive studies are more likely to be published compared to negative studies. Table 4 shows general consistency in the definition of 'published.' However, two empirical studies [3,27] considered grey literature in their definition of 'published' although information on full publications and grey literature publications are separated (Figures 5, 6). Although not considered in the definition of 'published', four empirical studies [26,[28][29][30] gave information on the grey literature or reports in preparation. Three empirical studies gave no information on their definition of 'published' [21,31,32]. In addition, results are presented for the percentage of studies not submitted for journal publication (7% to 58%), of studies submitted but not accepted for publication (0 to 20%) by the time of analysis of the cohort and the percentage of studies not published that were not submitted (63% to 100%). This implies that studies remain unpublished due largely to failure to submit rather than rejection by journals. The main findings of the empirical studies are shown in Table 5 and they are separated into study level and outcome level results. Eight of the included cohort studies [3,4,21,26,27,29,31,32] investigated results in relation to their statistical significance. One empirical study considered the importance of the results as rated by the investigator [30] and another empirical study considered confirmatory versus inconclusive results [29]. Five of the empirical studies [3,4,26,27,29] that examined the association between publication and statistical significance found that studies with statistically significant results were more likely to be published than those with non-significant results. Stern et al [4] reported that this finding was even stronger for their subgroup of clinical trials (Hazard Ratio (HR) 3.13 (95% confidence interval (CI) 1.76, 5.58), p = 0.0001) compared to all quantitative studies (HR 2.32 (95% CI 1.47, 3.66), p = 0.0003). One empirical study [32] found that studies with statistically significant results were more likely to be submitted for publication than those with non-significant results. Easterbrook et al [26] also found that study publication bias was greater with observational and laboratory-based experimental studies (Odds Ratio (OR) 3.79, 95% CI; 1.47, 9.76) than with RCTs (OR 0.84, 95% CI; 0.34, 2.09). However, two empirical studies [21,31] found no statistically significant evidence for study publication bias (RR 4 (95% CI 0.6, 32) p = 0.1 and OR 0.53 (95% CI 0.25, 1.1) p = 0.1).
Ioannidis et al [5] found that positive trials were submitted for publication more rapidly after completion than negative trials (median 1 vs 1.6 years, p,0.001) and were published more rapidly after submission (median 0.8 vs 1.1 years, p,0.04). Stern el al [4]  and Decullier et al [29] also considered time to publication and found that those studies with positive results were published faster than those with negative results (median 4.8 v 8.0 years [4] and HR 2.48 (95% CI 1.36, 4.55) [29], respectively).
Pich et al [28] looked at whether studies in their cohort were completed and published; 64% (92/143) of initiated trials were finished in accordance with the protocol and 31% (38/123) were published (or in-press) in peer reviewed journals.
Seven empirical studies [3,21,26,27,29,30,32] described reasons why a study was not published as reported by the trialists. Reasons related to trial results included: unimportant/ null results; results not interesting; results not statistically significant.
Outcome reporting bias. The total number of studies published in each cohort varied from 37% to 67% (Table 3). However, none of the empirical studies investigating ORB considered the proportions of published trials with positive, negative, or null overall results. Table 4 shows that three of the empirical studies [14,15,18] defined 'published' as a journal article; one empirical study [13] (Figure 15). Although not considered in the definition of 'published', one empirical study [14] gave information on the grey literature or reports in preparation. Only two empirical studies [14,15] present results for the percentage of studies not submitted (31% to 56%), the percentage of studies submitted but not accepted (1 to 2%) by the time of analysis of the cohort and the percentage of studies not published that were not submitted (97% to 99%).
All four empirical studies [14,15,17,18] that examined the association between outcome reporting bias (outcome level bias) and statistical significance found that statistically significant outcomes were more likely to be completely reported than nonsignificant outcomes (range of odds ratios: 2.2 to 4.7 (Table 5)).
Five empirical studies [13][14][15]17,18] compared the protocol and the publication with respect to the primary outcome (Table 5). Only two empirical studies looked at the different types of discrepancies that can arise [14,15] and concluded that 40-62% of trials had major discrepancies between the primary outcomes specified in protocols and those defined in the published articles. Four of the included empirical studies found that in 47-74% of studies the primary outcome stated in the protocol was the same as in the publication; between 13 and 31% of primary outcomes specified in the protocol were omitted in the publication and between 10 and 18% of reports introduced a primary outcome in the publication that was not specified in the protocol.
Chan et al also looked at efficacy and harm outcomes and in their Canadian empirical study [14] found that a median of 31% of efficacy outcomes and 59% of harm outcomes were incompletely reported and statistically significant efficacy outcomes had a higher odds than non significant efficacy outcomes of being fully reported (OR 2.7; 95% CI 1.5, 5). In their Danish empirical study [15] they found that 50% of efficacy and 65% of harm outcomes per trial were incompletely reported and statistically significant outcomes had a higher odds of being fully reported compared with non significant outcomes for both efficacy (OR 2.4, 95% CI; 1.4, 4) and harm (OR 4.7, 95% CI; 1.8, 12) data.
von Elm et al [18] considered efficacy and harm outcomes as well as primary outcomes overall and found that 32% (223/687) were reported in the publication but not specified in the protocol and 42% (227/546) were specified in the protocol but not reported, however this is preliminary data.
Two empirical studies [14,15] describe the reasons why outcomes do not get reported but the study is published, these include lack of clinical importance and lack of statistical significance.

Discussion
Very few empirical studies examined both study publication bias and outcome reporting bias in the same cohort. However, 12 of the included empirical studies demonstrate consistent evidence of an association between positive or statistically significant results and publication. They suggest that studies reporting positive/ statistically significant results are more likely to be published and that statistically significant outcomes have higher odds of being fully reported.
In this review we focused on empirical studies that included RCTs since they provide the best evidence of the efficacy of medical interventions [35]. RCTs are prone to study publication bias, but it has been shown that other types of studies are more prone to study publication bias [26]. The main limitation of this review was that for eight of the 16 included cohorts, information on RCTs could not be separated from information on other studies. Due to this barrier, and variability across empirical studies in the time lapse between when the protocol was approved and when the data were censored for analysis, we felt it was not appropriate to combine statistically the results from the different cohorts. Also, the fact that in five empirical studies [3,4,26,29,30] follow-up of trials was less than 90% could mean that the problem of study publication bias is underestimated in these cohorts.
It is difficult to tell the current state of the literature with respect to study publication bias, as even the most recently published empirical evaluations included in the review, considered RCTs which began 10 years ago. Nevertheless, the empirical studies that were published within the last eight years show that the total amount of studies published was less than 50% on average.
None of the empirical studies explored the idea of all outcomes being non-significant versus those deemed most important being non-significant. In the reasons given, it was not stated which outcomes/how many outcomes were non-significant. Some empirical studies imply that all results were non-significant although this is due to the way the reason was written i.e. no  [17]. doi:10.1371/journal.pone.0003081.g017 significant results; but it is not explained whether this means for all outcomes, or primary and secondary, harm and efficacy etc. This implies a potential ambiguity of 'no significant results'. It is not clear whether studies remain unpublished because all outcomes are non-significant and those that are published are so because significant results are selectively reported. This is where study publication bias and outcome reporting bias overlap.
Dubben et al [38] looked at whether study publication bias exists in studies which investigate the problem of study publication bias. Although they found no evidence of study publication bias, it is interesting to note that two of the included cohorts in this review have not been published [17,21]. The study conducted by Wormald et al [21] concluded that 'there was limited evidence of study publication bias' whereas the authors of the other study [17] have not as yet had time to submit the study for publication. The inclusion and exclusion criteria were applied by only the first author, and there may be other unpublished studies of study publication bias or outcome reporting bias that were not located by the search, however contact with experts in the field reduces the likelihood of these issues introducing bias. Submission is an important aspect of investigating study publication bias as it will provide information on whether reports are not being published because they are not submitted or they are submitted but not accepted. Obviously those studies that are not submitted are not published and it was found by Dickersin et al [36] that non-publication was primarily a result of failure to write up and submit the trial results rather than rejection of submitted manuscripts. This is confirmed for the cohorts identified here with the percentage of studies not published due to not being submitted ranging from 63% to 100%. Olson et al [37] also found that there was no evidence that study publication bias occurred once manuscripts had been submitted to a medical journal. However, this study looks at a high impact general journal, which is unlikely to be representative for specialist journals that publish the majority of clinical trials.
Ten studies assessed the impact of funding on publication; this was done in several ways. Three studies found that external funding lead to a higher rate of publication [4,27,30]. von Elm et al [18] found that the probability of publication decreased if the study was commercially funded and increased with non commercial funding. Easterbrook et al [26] found that compared with unfunded studies, government funded studies were more likely to yield statistically significant results but government sponsorship was not found to have a statistically significant effect on the likelihood of publication and company sponsored trials were less likely to be published or presented. Dickersin et al [3] found no difference in the funding mechanism grant versus contract and Ioannidis et al [5] found no difference in whether data was managed by the pharmaceutical industry or other federally sponsored organisations. Chan 2004b et al [15] found that 61% of the 51 trials with major discrepancies were funded solely by industry sources compared with 49% of the 51 trials without discrepancies. Ghersi [17] did examine the effect of funding in terms of reporting and discrepancies of outcomes but no information about the results is currently available. Hahn et al [13] compared the funder stated in protocol to publication. These studies indicate that funding is an important factor to consider when investigating publication bias and outcome reporting bias, however more work needs to be done to examine common questions before conclusions regarding the relationship between funding and outcome reporting bias can be drawn.
Our review has examined inception cohorts only, however, other authors have investigated aspects of study publication bias and outcome reporting bias using different study designs, with similar conclusions. The Cochrane review by Scherer et al [6] investigating the full publication of results initially presented in abstracts found that only 63% of results from abstracts describing randomized or controlled clinical trials are published in full and 'positive' results were more frequently published than non 'positive' results. Several studies investigated a cohort of trials submitted to drug licensing authorities [22,23,42] and all found that many of these trials remain unpublished, with one study demonstrating that trials with positive outcomes resulted more often in submission of a final report to the regulatory authority [22]. Olson et al [37] conducted a prospective cohort study of manuscripts submitted to JAMA and assessed whether the submitted manuscripts were more likely to be published if they reported positive results. They did not find a statistically significant difference in publication rates between those with positive and negative results. None of the inception cohorts addressed the question as to whether the significance determined whether a submitted paper was accepted or not, with the exception of one inception cohort [5] that found that ''positive'' trials were published significantly more rapidly after submission than ''negative'' trials. Finally, a comparison of the published version of RCTs in a specialist clinical journal with the original trial protocol found that important changes between protocol and published paper are common; the published primary outcome was   The study did not investigate this. 1 Gives results for trials with discrepancies for $1 primary outcome. 2 Results obtained from 248 of the completed studies. doi:10.1371/journal.pone.0003081.t005 Table 5. cont.
exactly the same as in the protocol in six out of 26 trials (23%) [43]. We recommend that researchers use the flow diagram presented in this work as the standard for reporting of future similar studies that look at study publication bias and ORB as it clearly shows what happens to all trials in the cohort.
Reviewers should scrutinise trials with missing outcome data and ensure that an attempt to contact trialists is always made if the study does not report results. Also, the lack of reporting of specified outcome(s) should not be an automatic reason for exclusion of studies. Statisticians should be involved for the data extraction of more complex outcomes, for example, time to event. Methods that have been developed to assess the robustness of the conclusions of systematic reviews to ORB [44,45] should be used. Meta-analyses of outcomes where several relevant trials have missing data should be seen with extra caution. In all, the credibility of clinical research findings may decrease when there is wide flexibility in the use of various outcomes and analysis in a specific field and this is coupled with selective reporting biases.
The setting up of clinical trials registers and the advance publication of detailed protocols with an explicit description of outcomes and analysis plans should help combat these problems. Trialists should be encouraged to describe legitimate changes to outcomes stated in the protocol. With the set up of online journals, where more space is available, trialists should be encouraged to write up and submit for publication without selection of results.
For empirical evaluations of selective reporting biases, the definition of significance is important as is whether the direction of the results is taken into account, i.e. whether the results are significant for or against the experimental intervention. However, only one study took this into account [5]. The selective publication preference forces may change over time. For example, it is often seen that initially studies favouring treatment are more likely to be published and those favouring control suppressed. However, as time passes, contradicting trials that favour control may become attractive for publication, as they are 'different.' The majority of cohorts included in this review do not consider this possibility.
Another recommendation is to conduct empirical evaluations looking at both ORB and study publication bias in RCTs to investigate the relative importance of both i.e. which type of bias is the greater problem. The effects of factors such as funding, i.e. the influence of pharmaceutical industry trials versus non pharmaceutical trials, should also be factored in these empirical evaluations.

Supporting Information
Appendix S1 Search Strategy.