This article has Open Peer Review reports available.
Assessment of regressionbased methods to adjust for publication bias through a comprehensive simulation study
 Santiago G Moreno^{1}Email author,
 Alex J Sutton^{1},
 AE Ades^{2},
 Tom D Stanley^{3},
 Keith R Abrams^{1},
 Jaime L Peters^{1} and
 Nicola J Cooper^{1}
https://doi.org/10.1186/1471228892
© Moreno et al; licensee BioMed Central Ltd. 2009
Received: 22 August 2008
Accepted: 12 January 2009
Published: 12 January 2009
Abstract
Background
In metaanalysis, the presence of funnel plot asymmetry is attributed to publication or other smallstudy effects, which causes larger effects to be observed in the smaller studies. This issue potentially mean inappropriate conclusions are drawn from a metaanalysis. If metaanalysis is to be used to inform decisionmaking, a reliable way to adjust pooled estimates for potential funnel plot asymmetry is required.
Methods
A comprehensive simulation study is presented to assess the performance of different adjustment methods including the novel application of several regressionbased methods (which are commonly applied to detect publication bias rather than adjust for it) and the popular Trim & Fill algorithm. Metaanalyses with binary outcomes, analysed on the log odds ratio scale, were simulated by considering scenarios with and without i) publication bias and; ii) heterogeneity. Publication bias was induced through two underlying mechanisms assuming the probability of publication depends on i) the study effect size; or ii) the pvalue.
Results
The performance of all methods tended to worsen as unexplained heterogeneity increased and the number of studies in the metaanalysis decreased. Applying the methods conditional on an initial test for the presence of funnel plot asymmetry generally provided poorer performance than the unconditional use of the adjustment method. Several of the regression based methods consistently outperformed the Trim & Fill estimators.
Conclusion
Regressionbased adjustments for publication bias and other small study effects are easy to conduct and outperformed more established methods over a wide range of simulation scenarios.
Background
Publication bias (PB) has the potential to distort the scientific literature [1, 2]; since it is the "interest level", or statistical significance of findings, not study rigour or quality, that determines which research gets published and is subsequently available [3]. A metaanalysis of the published literature will be biased and may adversely affect decision making if PB exists.
More generally, the tendency for smaller studies to show a greater effect than larger studies, when evaluating interventions, has been named smallstudy effects [4]. These could be due to publication bias, or other factors. Any factor which confounds the relationship betweenstudy effect and study size may cause smallstudy effects. For example, if an intervention is more effective in highrisk patients, and the small studies are, on average, conducted in higherrisk patients, this may result in a larger treatment efficacy being observed in the smaller studies. Further, it has been observed that certain aspects of trial quality influence effect size estimates and empirical evidence suggests that small studies are, on average, of lower quality [4].
Several statistical methods exist for detecting funnel plot asymmetry/smallstudy effects [6]. However, detection alone is i) limited since the likely impact of the bias is not assessed, [7]; ii) problematic since the chance of a false negative result are typically high; [8] and iii) insufficient if the results of the metaanalysis are to inform policy decisions. A reliable way to adjust metaanalysis for smallstudy effects is required to facilitate more reliable decisionmaking.
Disentangling the underlying cause of funnel plot asymmetry is difficult, although adding contours of statistical significance onto the plot has recently been suggested as a way of aiding interpretation [9]. However, whether the cause for funnel asymmetry is publication bias or other factors, predicting the effect in an infinitely large study – at the top of a funnel plot – can be perceived to be unbiased since i) if publication bias is the cause, larger studies will be less affected than smaller ones under the selection mechanisms assumed to underlie publication bias [10] (i.e. a hypothetical study of infinite size would have no chance of being suppressed and hence would provide an unbiased estimate of the population effect); and ii) if other confounding factors are the source of the smallstudy effects, the effect sizes of larger studies better represent the effects that would be seen when an intervention is implemented on a large scale, while smaller studies tend to be less representative of the population of interest. (Again, the effect size of interest is best represented by a hypothetical study of infinite size.)
We evaluate, through a comprehensive simulation study, a number of regressionbased approaches to adjust a metaanalysis for publication bias. We also evaluate the performance of the models conditional on a statistically significant result from a test for publication bias. For comparison we also consider the Trim & Fill method [11], which is probably the most widely used adjustment method presently.
In the simulation study we have generated smallstudy effects by inducing PB using different models for study suppression. Therefore, for simplicity, we refer to PB as the cause of such smallstudy effects in the remainder of this paper. However, as argued above, we consider such an approach to be appropriate for all smallstudy effects in a decision making context.
The outline of this article is as follows. Section two presents a description of the statistical methods assessed. Section three describes the design of the simulation study. Section four presents the results of the simulation study, and Section five, the discussion, concludes the article.
Methods
Adjustment methods evaluated
The context considered throughout is that of 2arm comparative studies reporting binary outcome data, with the metaanalysis being conducted on the (log) odds ratio scale. The different regressionbased methods used to adjust for PB are described below. Additionally, for completeness we also consider the standard fixed and random effects metaanalysis models [12].
Trim and fill
Trim & Fill [11, 13] is probably the most popular method for examining the possible effect of PB on the pooled estimate, and can be defined as an iterative nonparametric adjustment method based on a rankbased data augmentation technique to account for asymmetry on the funnel plot. Briefly, the method "trims" the asymmetric studies on the righthand side of the funnel for which there are no lefthand counterparts. A revised pooled estimate "adjusting for publication bias" is then derived from this reduced dataset. Then, the "trimmed" studies are reinstated and studies, assumed to be missing, are imputed on the opposite side of the funnel by "reflecting" the trimmed studies about the adjusted pooled effect line and uncertainty in the "adjusted" pooled effect is calculated using this augmented dataset.
We evaluate both L_{0} and R_{0} estimators [13] in the simulation study. However, since results were similar for both estimators relative to differences with the other methods, only the results from the R_{0} estimator are presented. We implement this method using fixed effects in both the 'trimming' and 'filling' parts of the algorithm (fixedfixed), similarly, random effects for both parts (randomrandom), and fixed effects to 'trim' and random effects to 'fill' (fixedrandom). Justification and evaluation of these variants is available elsewhere. [6, 14]
Regression methods
A number of regression models exist that assess the degree of association between the study effect and a measure of its precision. The initial test based on the statistical significance of this association was suggested by Egger et al. [10], but due to concerns regarding its performance on the odds ratio scale, several modifications have now been proposed and are also considered [15, 16]. A further arcsinebased regression test developed by Rücker et al. [17] is not considered further here because it performs a correction on the arcsine scale, which is harder to directly compare with the other methods performance.
Evidence of such an association may suggest that the metaanalysis is affected by PB(/smallstudy effects) if the observed smaller, less precise, studies have larger effect sizes than the more precise studies [4, 18]. This association can be illustrated by the regression line on a funnel plot (figure 1). This figure presents a standard funnel plot [19] of outcome (ln(OR)) against a measure of study precision (se(lnOR)), for a simulated dataset with an underlying ln(OR) of 0.4 but with PB induced (i.e. studies are suppressed in bottom left hand side of plot). This regression line indicates how the less precise studies tend to have, on average, larger treatment effects, implying that PB(/small study effects) bias exist.
By extrapolating this regression line to the point where the standard error is zero, the effect size from a (hypothetical) infinitely large study can be predicted. And, as discussed in Section 1, this can be interpreted as the effect size adjusted for PB(/smallstudy effects). For example in Figure 1, a standard fixed effects metaanalysis estimates a pooled effect size of ln(OR) = 0.58, which is considerably larger than the true underlying effect (lnOR = 0.4). When metaregression is applied to the dataset, where the independent variable is se(ln(OR)), this predicts a ln(OR) of 0.38 for a se(ln(OR)) of 0, which is closer to the underlying truth. The notion of adjusting for PB through incorporation of study precision as the metaregression covariate has been suggested previously [4, 20–22] but not formally evaluated. Details of these various regression methods applied in this way are described below.
where y _{ i }is the ln(OR) from study i (in the context considered in this paper) and se _{ i }is its associated standard error. We interpret the two coefficients α and β to represent the adjusted pooled effect (intercept) and the slope associated with funnel plot asymmetry respectively. The regression is weighted by the inverse of the estimated effect size variance for each study (1/se _{ i } ^{2}). φ is an unknown multiplicative dispersion parameter that is estimated in the regression and allows for possible heteroscedasticity [23, 24].
Here, μ _{ i }is a normal error term with mean zero and variance τ ^{2} to be estimated from the data.
where Z _{ i }is the efficient score and V _{ i }is Fisher's information (the variance of Z under the null hypothesis for the i ^{ th }study).
As before, the two coefficients α and β represent the adjusted pooled effect (intercept) and the regression slope respectively. For each study i, a and b represent the observed number who experience the outcome of interest in the treated and control groups, respectively, and c and d are the numbers corresponding to those not developing the outcome in the treated and control group respectively. Thus, the sample size of the i ^{ th }study corresponds to the sum of a _{ i } , b _{ i } , c _{ i }and d _{ i }.
Conditional methods
We suspect, in practice, researchers carry out a test for smallstudy effects and consider the use of adjustment methods conditional on the outcome of such a test. Therefore, we also evaluate two conditional approaches in which a standard random effects model or either of the original Egger or Harbord adjustment based methods are used depending on whether the corresponding test (i.e. Egger or Harbord respectively) was significant at the 10% level. Since the Egger conditional approach is almost always outperformed by the Harbord conditional method, only the latter is reported below.
Summary of adjustment methods
In summary, the performance of the following adjustment methods is reported below. An abbreviation is given to each method, which is used in the remainder of the paper:

The two usual standard metaanalysis methods

Fixed effects metaanalysis (FE)

; Random effects metaanalysis (RE)


Nonparametric adjustment method: Trim & Fill

R_{0} estimator, trim using fixed effects & fixed effects on filled dataset (TF FEFE)

R_{0} estimator, trim using fixed effects & random effects on filled dataset (TF FERE)

R_{0} estimator, trim using random effects & random effects on filled dataset (TF RERE)


Parametric adjustment methods: weighted regressions

Egger's model variants:

Fixed effects (FEse);

Random effects (REse)

Dispersion (Dse)


EggerVar model variants:

Fixed effects (FEvar)

Random effects (REvar)

Dispersion (Dvar)


Other regressions

Harbord's model (Harbord)

Peters' model (Peters)



Conditional method

PB test plus conditional adjustment based on it using Harbord's model (HarbordC)

Simulation Study
Methods for generating the metaanalysis datasets
Simulated metaanalyses were based on a set of characteristics intended to reflect metaanalyses of randomised clinical trials in the medical literature. The assumptions made and parameter values chosen have drawn on the authors' extensive experience in this area as well as considering the complete review of previous simulation studies in the field [27].
Scenarios in which 5, 10, 20 or 30 individual trials were included in the metaanalysis were explored [4]. The sample size of the individual studies within each metaanalysis was generated from a log normal distribution with mean 6 and variance 0.6. This reflects the greater number of small studies compared to large studies as commonly observed in real metaanalyses. This distribution results in a mean (median) size of 483 (403) individuals per study and a standard deviation of 318. The 1% (99%) percentile is 100 (1628) individuals per study. The numbers of individuals allocated to treatment and control arms was equal for all simulations.
Both fixed (homogeneous) and random effects (heterogeneous) metaanalysis scenarios were simulated. Underlying effect sizes (i.e. for all studies under fixed effects and for the mean of the distribution of studies under random effects) considered were OR = 1 (null effect), 1.5 and 3 (representing a large effect), where OR > 1 is considered clinically beneficial. Following the approach of Schwarzer's et al. [28], the event rate for the intervention and control arms is modelled by simulating the average event probability of the treatment and control trial arms. (This approach reduces the correlation between individual effect estimates and their corresponding standard errors compared to modelling the event rate on the control group). The average event probability for each trial was generated according to a uniform distribution (0.3, 0.7). The actual number of events in each study arm was generated according to a binomial distribution taking into account the corresponding arm event probability and study arm size.
Where underlying ln(OR) is δ, and μ _{ i }is the average event rate on the logit scale in the i ^{ th }study. n _{ i } , p _{ i } r _{ i }are the number of subjects, the probability of an event (derived directly from μ _{ i }and δ), and the number of events in the i ^{ th }study arm respectively, with a superscript C or T indicating the control or treatment group.
Now the underlying effect in the i ^{ th }trial, δ _{ i }is assumed to be drawn from a Normal distribution with mean value θ and betweenstudy variance τ ^{2}. τ ^{2} is defined to be either 0%, 100%, 150% or 200% of the average withinstudy variance for studies from the corresponding fixed effects metaanalyses simulated. [15]
The betweenstudy variance can also be defined in terms of I ^{2} (the percentage of total variation across studies that is due to betweenstudy variation rather than sampling error [32]). In a scenario where PB is absent, a betweenstudy variation of 0% & 150% of the average withinstudy variation corresponds to an average I ^{2} of 7% & 57% respectively. Note that under scenarios where PB is simulated, this will affect estimates of the betweenstudy variability [33].
The two most commonly assumed selection processes used in previous simulation studies to induce PB are considered here:
1) Publication suppression on the basis of a onesided pvalue associated with the effect estimate of interest [14, 25, 33–36]
Specification of publication bias severity based on onesided significance
Severity of Publication Bias  pvalue from study  Probability study included in MA 

Moderate  < 0.05  1 
0.05 – 0.5  0.75  
> 0.5  0.25  
Severe  < 0.05  1 
0.05 – 0.2  0.75  
> 0.2  0.25 
2) Suppressing the most extreme unfavourable results
This assumes that only the estimated effect size influences whether a study is included in the metaanalysis or not, so that studies with the most extreme unfavourable estimates of effect are excluded [11, 14]. Here, the number of studies excluded does not depend on the underlying effect size, as it does when PB is induced on the basis of pvalue. "Moderate" and "severe" PB were represented by excluding either the 14% or 30% of the most extreme studies showing an unfavourable effect such that the final number of studies in a metaanalysis was reached. For example, under severe level of PB, where the metaanalysis size ought to be 30, 50 studies are generated so that 20 studies (i.e. 30% of the original 50 studies) giving the most extreme unfavourable estimates are omitted from the metaanalysis.
Simulation scenarios investigated
Altogether there are 5 different PB situations (none, and moderate and severe for both pvalue and effect size suppression). Each of these situations is applied to all combinations of the following metaanalysis characteristics: underlying effect size (OR = 1, 1.5 & 3), number of studies in the metaanalysis (5, 10, 20 & 30) and degree of betweenstudy heterogeneity (0%, 100%, 150% or 200% of the average withinstudy variance). This results in 240 individual scenarios, with 5,000 simulated datasets generated for each scenario. All methods summarised in section 2.4 were applied to each of these datasets. All statistical analyses were performed using Stata, version 9.2 [38].
Results of this simulation study are reported in section 4; although not all scenarios are shown due to the negligible added value they provide to the overall conclusions. In particular, scenarios with "moderate" bias are omitted because they follow the same overall trend as the more severe scenarios, but with differences between methods tending to be less pronounced. Hence the PB situations labelled 1, 2 and 4 in Figure 2 (none and severe by both mechanisms) are reported here. (For completeness, a webonly appendix containing the remaining scenarios are available from [see Additional file 1]).
Criteria to assess the methods performance
In order to evaluate the adjustment methods, their performance is assessed through measures of model accuracy (bias), precision (variance) and coverage (type I error rate) [39] in the simulation study.
We consider "absolute bias" – the expected difference between the estimated (adjusted) effect and the underlying true effect across all simulations. A negative bias indicates an underestimate of the true underlying effect, and a positive residual bias indicates an overestimate of the true underlying effect. Arguably, when interpreting absolute bias, it is important to consider how close the true effect is from the null effect. It is desirable for the adjustment methods to work well near the null where small changes may have impact on the direction of the effect and hence any conclusions. Conversely, if the performance is poorer further away from the null effect (e.g. OR = 3), the consequences are, in some respects, less of a concern because it is unlikely that PB will change the direction of the pooled effect.
Where ${\widehat{\theta}}_{j}$ is the estimated ln(OR) predicted by the model for the j ^{th} simulated metaanalysis (j:1, ..., N; where N is the total number of simulations). The underlying true value of ${\widehat{\theta}}_{j}$ is θ. The MSE can be thought of as corresponding to the sum of the variance plus the square of bias of $\widehat{\theta}$.
We also consider the coverage probability, which can be defined as the proportion of simulations in which the true underlying effect lies within the 95% confidence interval of the predicted effects simulated. This informs how well the type I error is controlled by the statistical model. The coverage probabilities should be approximately equal to the nominal coverage rate to properly control the type I error rate for testing a null hypothesis of no difference in effect size between the true underlying effect and the predicted one; the 5% level is used throughout.
The final measure considered is the average "variance" of the predicted pooled effects. This is used to ascertain the contribution of bias and variance to the MSE.
Results
When the underlying effect is increased to OR = 3 (figure 4), greater variability in results of the different methods is observed, with generally worse performance seen for most methods. This is due, at least in part, to the susceptibility of some methods to the induced artefactual relationship between ln(OR) and its se(lnOR) which increases as the OR increases. Indeed, while the standard random effect model has no bias, the two adjustment methods with lowest absolute bias are those using the Peters and the conditional Harbord models which were developed to circumvent the problems of the artefactual relationship. Note also that the Harbord, Egger RE (REse) and dispersion (Dse) methods report large MSE values. This can be partially explained by their large variance in the model predictor. With respect to the Egger based methods, the use of the variance as the predictor variable provides less biased and more precise estimates than using the standard error.
The degree of absolute bias is dependent on the underlying odds ratio. This can be explained by pvalue induced PB causing 'disfigurement' to the funnel plot which is dependent on the underlying odds ratio; i.e. the funnel plot is almost intact under OR = 3, while a very asymmetrical shape is obtained under OR = 1. The methods that can accommodate heterogeneous data (through the inclusion of random effects or dispersion parameters) are the ones with the most appropriate coverage probabilities among the ones evaluated; these include the Harbord, Peters as well as the Eggerbased methods (REse, Dse, REvar & Dvar). However, the Harbord and the two Egger (REse & Dse) methods report substantially inflated MSE and variance values compared to other methods evaluated.
The conditional regression method, Trim & Fill and standard metaanalysis estimators do not perform particularly well due to low coverage and large amounts of residual bias. As before, for larger values of heterogeneity, the EggerVar methods (FEvar, REvar & Dvar) tend to report somewhat lower MSE and coverage probabilities than the Egger ones (FEse, REse & Dse) thanks to their restrained variance.
Also of note is that, under fairly homogeneous effects, adjustment methods based on FE (FEse & FEvar) and RE (REse & REvar) Egger models provide coverage probabilities well above 95%, implying inappropriately small type I error rates. This can be explained by the inability of these models to accommodate the underdispersion of observed effects (i.e. less variability than would be expected by chance) caused by PB. Conversely, the methods which include a dispersion parameter do not suffer from excessive coverage probabilities due to the fact that they can accommodate underdispersion by allowing the dispersion parameter below the value of one. This means that the Harbord, Peters and the Egger methods (Dse & Dvar) perform favourably under scenarios where PB causes underdispersion.
Discussion
In this paper we have compared some novel and existing methods for adjusting for publication bias through an extensive simulation study. Results are encouraging, with several of the regression methods displaying good performance profiles. Overall, no particular method consistently outperforms all others. The overall performance of all the methods deteriorates as I ^{2} exceeds 50% [30] and the underlying odds ratio increase; while at the same time differences between them diverge.
With respect to the popular Trim & Fill method, we find it hard to recommend over the regressionbased alternatives due to its potentially misleading adjustments and poor coverage probabilities, especially when betweenstudy variance is present [14, 29], although it should be acknowledged that Trim & Fill was only intended as a form of sensitivity analysis [11] rather than as an adjustment method per se.
Although the standard metaanalysis models are a good approach under lack of PB, they inevitably perform poorly when PB is present. This motivated the examination of regressionbased adjustment methods conditional on their associated test for PB. Such an approach is also of interest because it may reflect what is commonly done in practice when dealing with suspected PB. Unfortunately, these conditional approaches did not perform as well as the (unconditional) alternatives. This may be explained by the fact that all existing tests for PB suffer from low statistical power [4, 7, 25] leading to inappropriate methods being used in some instances, and, this is a warning to not using such an approach (formal or informally). This is an inherent problem of pretests since the failure of the pretest to reject the nullhypothesis does not prove the null hypothesis true, unless the pretest was designed as an equivalence test.
The persistent low level of coverage probability by the fixed effects Egger models (FEse & FEvar) under heterogeneous settings render them inappropriate. Equally, coverage probabilities above the 95% threshold produce inaccuracy on the confidence interval; which potentially biases any subsequent assessment of uncertainty around the estimate of interest. This is a serious concern in a decisionmaking context, where alternative treatments may report similar mean effect sizes. In such cases, accurate quantification of uncertainty to allow discriminating among treatments is vital to facilitate realistic probabilistic statements about, say, costeffectiveness relative to the alternative treatments. Here, both fixed effects (FEse & FEvar) and random effects (REse & REvar) Egger models tend to suffer from excessive coverage probabilities under:
• Scenarios of underdispersion caused by severe PB (figure 8);
• Mostly homogeneous settings (figures 5, 7, 8), provided the metaanalysis is not exceptionally large (i.e. less than 30 studies); and
• Small size metaanalysis (figure 5), provided the data is fairly homogeneous.
Additionally, since in practice it will often be difficult to determine whether heterogeneity is present or not (due to the low power of associated test and distortions caused by PB) this makes appropriate implementation of fixed effect methods difficult.
Over the range of simulation scenarios considered, the Harbord, Peters and both Egger dispersion (Dse & Dvar) methods would appear to have best overall performance. They do not always produce the least biased estimate, but they do consistently retain good coverage probability levels (by equally accommodating homogeneous and heterogeneous data), while keeping competitive with respect to bias.
However, when faced with small size metaanalyses and/or heterogeneity (figures 3, 4, 5, 6, 7, 8), the outstanding coverage comes to a high cost in terms of MSE for the Harbord and Egger dispersion (Dse) methods compared to the other two. These two methods tend to report low residual bias but yet persistent high MSE values, due to the large variances. In contrast, Peters and the Eggervar (Dvar) methods report slightly lower coverage probabilities besides much lower MSE values as a result of their restrained variance. Due to this, we recommend the Peters and Eggervar (Dvar) methods which perform very similarly throughout the simulations: at least in terms of coverage, MSE and variance. However, there is one instance (figure 4) where they clearly differ with regard to absolute bias; which can be explained by the Peters' method profiting from avoiding the structural correlation problem between outcome and standard error by using a function of sample size as the predictor variable.
One favourable factor in this simulation study is that there was always considerable variation in the sizes of the studies in each dataset. Again, the methods performance will deteriorate if studies sizes are less variable. This is particularly a concern for the regression approaches if all the studies are small, since a larger extrapolation to the intercept would be required.
In these simulations we defined levels of heterogeneity in terms if I ^{2} (the percentage of total variation across studies that is due to betweenstudy variation rather than sampling error). By doing this, heterogeneity is induced proportionally to the withinstudy variation. By defining heterogeneity in terms of the I ^{2} statistic means we are focussing on the impact rather than the extent of heterogeneity [32] across the different metaanalytic scenarios. An alternative modelling approach would be to define heterogeneity in terms of the betweenstudy variance parameter (τ ^{2}) which would lead to an assessment of the methods with respect to absolute degrees of betweenstudy variability. Previous studies that evaluate publication bias methods [15–17, 28] have used a mixture of these approaches and it is not clear which, if either, is superior.
Other methods for PB adjustment are available but were not evaluated in the simulation study. These include a literature on the use of selection modelling techniques [40]. The reason for excluding them is twofold: 1) Unless there are large numbers of studies, it will be necessary to specify the selection mechanism as a modelling assumption. Hence their performance will directly depend on how good the specification of the selection model is and this is difficult to evaluate via simulation (i.e. if you specify the selection model to be the same as used to simulate the data you can guarantee good performance and vice versa). 2) Previous work has acknowledged that since the selection mechanism is not identifiable from the data, sensitivity analyses should be carried out using a range of selection functions. While this is potentially useful in an inference making context where robustness or lack of it may be explored over a range of possible selection models, it is less useful in a decision making context where a single decision has to be made.
Recently Copas and Malley [22] presented a novel way of obtaining a robust pvalue for effect in a metaanalysis with publication bias based on a permutation test. Interestingly, this is shown to be closely related to the correlation found in the associated radial plot, which in turn is closely related to a funnelplot related regression [4].
Since in medical applications any PB selection mechanisms will be unknown and there will often be too few studies to estimate it from the data, we believe regressionbased methods, which make no explicit assumptions about the underlying selection mechanism, may have a useful role in a decisionbased context.
We believe that a broad range of plausible metaanalyses situations have been evaluated through the scenarios evaluated in the simulation study. And that, given the variability and limited scope of some of the previous simulation studies in the evaluation of methods to address PB in the past, it would be desirable for there to be a consensus simulation framework in which future tests and adjustment methods could be evaluated. To this end, the comprehensive framework developed here could form the starting point for future simulation studies.
Conclusion
In conclusion, several regressionbased models for PB adjustment performed better than either the Trim & Fill or conditional (regressionbased) approaches. Overall, the Eggervar (Dvar) and Peters methods are identified as methods with potentially appealing statistical properties for PB adjustment with the Peters method performing better for large odds ratios under the simulation scenarios evaluated. However, it should be acknowledged that while our simulations were extensive, differences in results may be observed if different simulation parameter values were used.
Further research is considered worthwhile given our encouraging initial results. To this end, further work exploring the incorporation of information obtained from external sources to form a prior distribution for the regression coefficients is being developed in the hope of improving performance of the regressionbased methods.
Finally, while we acknowledge that while publication bias is a problem that will not entirely disappear regardless of the statistical method of analysis, ignoring it is an unwise option [41]. We also support prevention of PB as a more desirable approach compared to detection or adjustment [8, 16, 42]. However, despite the limitations of existing methods we believe it is helpful to attempt to adjust for PB as long as it is present in the literature, particularly in a decision making context [34].
Declarations
Acknowledgements
The authors would like to thank John Thompson and Tom Palmer for useful comments and discussions regarding the work contained in this paper. They would also like to thank Roger Harbord for discussions relating to the regression model he jointly developed.
Authors’ Affiliations
References
 Turner EH, Matthews AM, Linardatos E, Tell RA, Rosenthal R: Selective Publication of Antidepressant Trials and Its Influence on Apparent Efficacy. N Engl J Med. 2008, 358 (3): 25210.1056/NEJMsa065779.View ArticlePubMedGoogle Scholar
 Metcalfe S, Burgess C, Laking G, Evans J, Wells S, Crausaz S: Trastuzumab: possible publication bias. Lancet. 2008, 371 (9625): 16461648. 10.1016/S01406736(08)607060.View ArticlePubMedGoogle Scholar
 Smith ML: Publication bias and metaanalysis. Evaluation in Education. 1980, 4: 2224. 10.1016/0191765X(80)90004X.View ArticleGoogle Scholar
 Sterne JAC, Gavaghan D, Egger M: Publication and related bias in metaanalysis: Power of statistical tests and prevalence in the literature. J Clin Epidemiol. 2000, 53: 11191129. 10.1016/S08954356(00)002420.View ArticlePubMedGoogle Scholar
 Sterne JAC, Egger M: Funnel plots for detecting bias in metaanalysis: Guidelines on choice of axis. J Clin Epidemiol. 2001, 54: 10461055. 10.1016/S08954356(01)003778.View ArticlePubMedGoogle Scholar
 Rothstein HR, Sutton AJ, Borenstein M: Publication bias in metaanalysis. prevention, assessment and adjustments. 2005, Chichester: WileyView ArticleGoogle Scholar
 Ioannidis JPA, Trikalinos TA: The appropriateness of asymmetry tests for publication bias in metaanalyses: a large survey. CMAJ. 2007, 176 (8): 10911096.View ArticlePubMedPubMed CentralGoogle Scholar
 Lau J, Ioannidis J, Terrin N, Schmid C, Olkin I: The case of the misleading funnel plot. BMJ. 2006, 333 (7568): 597600. 10.1136/bmj.333.7568.597.View ArticlePubMedPubMed CentralGoogle Scholar
 Peters J, Sutton AJ, Jones DR, Abrams KR, Rushton L: Contourenhanced metaanalysis funnel plots help distinguish publication bias from other causes of asymmetry. J Clin Epidemiol. 2008, 991996. 10.1016/j.jclinepi.2007.11.010. 61Google Scholar
 Egger M, Smith GD, Schneider M, Minder C: Bias in metaanalysis detected by a simple, graphical test. BMJ. 1997, 315: 629634.View ArticlePubMedPubMed CentralGoogle Scholar
 Duval S, Tweedie RL: Trim and fill: A simple funnel plot based method of testing and adjusting for publication bias in metaanalysis. Biometrics. 2000, 56: 455463. 10.1111/j.0006341X.2000.00455.x.View ArticlePubMedGoogle Scholar
 Sutton AJ, Abrams KR, Jones DR, Sheldon TA, Song F: Methods for metaanalysis in medical research. 2000, John Wiley: LondonGoogle Scholar
 Duval S, Tweedie RL: A Nonparametric "Trim and Fill" Method of Accounting for Publication Bias in MetaAnalysis. J Am Stat Assoc. 2000, 95 (449): 8998. 10.2307/2669529.Google Scholar
 Peters JL, Sutton JA, Jones DR, Abrams KR, Rushton L: Performance of the trim and fill method in the presence of publication bias and betweenstudy heterogeneity. Stat Med. 2007, 26 (25): 45444562. 10.1002/sim.2889.View ArticlePubMedGoogle Scholar
 Peters JL, Sutton AJ, Jones DR, Abrams KR, Rushton L: Comparison of two methods to detect publication bias in metaanalysis. JAMA. 2006, 295 (6): 676680. 10.1001/jama.295.6.676.View ArticlePubMedGoogle Scholar
 Harbord RM, Egger M, Sterne JAC: A modified test for smallstudy effects in metaanalyses of controlled trials with binary endpoints. Stat Med. 2006, 25 (20): 34433457. 10.1002/sim.2380.View ArticlePubMedGoogle Scholar
 Rucker G, Schwarzer G, Carpenter J: Arcsine test for publication bias in metaanalyses with binary outcomes. Stat Med. 2008, 27 (5): 746763. 10.1002/sim.2971.View ArticlePubMedGoogle Scholar
 Egger M, Smith GD: Misleading metaanalysis [editorial]. BMJ. 1995, 310: 752754.View ArticlePubMedPubMed CentralGoogle Scholar
 Light RJ, Pillemar DB: Summing up: the science of reviewing research. 1984, Harvard University Press: Cambridge, MassGoogle Scholar
 Steichen TJ, Egger M, Sterne J: Tests for publication bias in metaanalysis. Stata Tech Bull. 1998, 44 (sbe20:915): 915.Google Scholar
 Stangl DK, Berry DA: 2000, MetaAnalysis in Medicine and Health Policy: Routledge, USAGoogle Scholar
 Copas JB, Malley PF: A robust Pvalue for treatment effect in metaanalysis with publication bias. Stat Med. 2008, 27 (21): 42674278. 10.1002/sim.3284.View ArticlePubMedGoogle Scholar
 Thompson SG, Sharp SJ: Explaining heterogeneity in metaanalysis: a comparison of methods. Stat Med. 1999, 18: 26932708. 10.1002/(SICI)10970258(19991030)18:20<2693::AIDSIM235>3.0.CO;2V.View ArticlePubMedGoogle Scholar
 McCullagh P, Nelder J: Generalized linear models. 1989, London: Chapman and HallView ArticleGoogle Scholar
 Macaskill P, Walter SD, Irwig L: A comparison of methods to detect publication bias in metaanalysis. Stat Med. 2001, 20: 641654. 10.1002/sim.698.View ArticlePubMedGoogle Scholar
 Hardy RJ, Thompson SG: A likelihood approach to metaanalysis with random effects. Stat Med. 1996, 15: 619629. 10.1002/(SICI)10970258(19960330)15:6<619::AIDSIM188>3.0.CO;2A.View ArticlePubMedGoogle Scholar
 Peters JL, Sutton AJ, Jones DR, Abrams KR: Performance of Tests and Adjustments for Publication Bias in the Presence of Heterogeneity. Technical report 0501. 2005, Department of Health Sciences, University of LeicesterGoogle Scholar
 Schwarzer G, Antes G, Schumacher M: Inflation of type I error rate in two statistical tests for the detection of publication bias in metaanalyses with binary outcomes. Stat Med. 2002, 21 (17): 24652477. 10.1002/sim.1224.View ArticlePubMedGoogle Scholar
 Terrin N, Schmid CH, Lau J, Olkin I: Adjusting for publication bias in the presence of heterogeneity. Stat Med. 2003, 22: 21132126. 10.1002/sim.1461.View ArticlePubMedGoogle Scholar
 Stanley TD: Metaregression methods for detecting and estimating empirical effects in the presence of publication selection. Oxford Bulletin of Economics and Statistics. 2008, 70: 103127.Google Scholar
 Lu G, Ades AE: Combination of direct and indirect evidence in mixed treatment comparisons. Stat Med. 2004, 23 (20): 31053124. 10.1002/sim.1875.View ArticlePubMedGoogle Scholar
 Higgins JPT, Thompson SG: Quantifying heterogeneity in metaanalysis. Stat Med. 2002, 21: 15391558. 10.1002/sim.1186.View ArticlePubMedGoogle Scholar
 Jackson D: The implications of publication bias for metaanalysis' other parameter. Stat Med. 2006, 25 (17): 29112921. 10.1002/sim.2293.View ArticlePubMedGoogle Scholar
 Hedges LV, Vevea JL: Estimating effects size under publication bias: small sample properties and robustness of a random effects selection model. Journal of Educational and Behavioural Statistics. 1996, 21 (4): 299332.View ArticleGoogle Scholar
 Preston C, Ashby D, Smyth R: Adjusting for publication bias: modelling the selection process. J Eval Clin Pract. 2004, 10 (2): 313322. 10.1111/j.13652753.2003.00457.x.View ArticlePubMedGoogle Scholar
 Copas J: What works?: selectivity models and metaanalysis. Journal of the Royal Statistical Society, Series A. 1998, 161: 95105.View ArticleGoogle Scholar
 Peters JL: Generalised synthesis methods in human health risk assessment. PhD thesis. 2006, University of Leicester, Department of Health SciencesGoogle Scholar
 Stata: Stata Statistical Software: Release 9.2. College Station, Texas. 2008Google Scholar
 Burton A, Altman DG, Royston P, Holder RL: The design of simulation studies in medical statistics. Stat Med. 2006, 25 (24): 42794292. 10.1002/sim.2673.View ArticlePubMedGoogle Scholar
 Vevea JL, Woods CM: Publication bias in research synthesis: sensitivity analysis using a priori weight functions. Psychol Methods. 2005, 10 (4): 428443. 10.1037/1082989X.10.4.428.View ArticlePubMedGoogle Scholar
 Baker R, Jackson D: Using Journal Impact Factors to Correct for the Publication Bias of Medical Studies. Biometrics. 2006, 62 (3): 785792. 10.1111/j.15410420.2005.00513.x.View ArticlePubMedGoogle Scholar
 Begg CB, Berlin JA: Publication bias: a problem in interpreting medical data (with discussion). Journal of the Royal Statistical Society, Series B. 1988, 151: 419463.View ArticleGoogle Scholar
 The prepublication history for this paper can be accessed here:http://www.biomedcentral.com/14712288/9/2/prepub
Prepublication history
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Comments
View archived comments (1)