You are on page 1of 4

STATISTICS AND RESEARCH DESIGN

Publication bias: Graphical and


statistical methods
Loukia M. Spinelia and Nikolaos Pandisb
Hannover, Germany, and Bern, Switzerland

V
arious statistical approaches and visual tools much wider the confidence interval under the random-
have been developed to detect, estimate, and effects model is compared with the confidence interval
evaluate the impact of publication bias in under the fixed-effect model. The similarity of the results
meta-analysis results. In this article, we present the is an indication of the possible low impact of small-study
most popular statistical methods and graphic tools to effects. Forest plots provide only a visual exploration;
address publication bias using an example. hence, further investigation is required (eg, funnel plot
and proper statistical methods) to determine any
small-study effects and possible publication bias.
COMPARING FIXED-EFFECT AND RANDOM-
EFFECTS RESULTS
The fixed-effect model is known to favor larger FUNNEL PLOT FOR SMALL-STUDY EFFECTS AND
studies; hence, it assigns greater weights to these PUBLICATION BIAS
studies. By contrast, the random-effects model aims to Another graphical tool to investigate the relationship
balance the weights more evenly across small and large between study size and effect size is the funnel plot.4
studies. With substantial small-study effects, where an The funnel plot is a scatter plot in which the effect sizes
intervention seems to be more beneficial in smaller are plotted on the x-axis and the standard errors of the
studies, the random-effects summary effect size will pre- effect sizes on the y-axis. The spread of the points cre-
sent the intervention as being more beneficial than the ates a pattern like a funnel. In the funnel plot, the points
fixed-effect summary effect size.1 corresponding to studies with smaller sample size are
The small-study effects issue is one of many factors scattered on the bottom of the funnel (because they
responsible for heterogeneity in the meta-analysis re- yield effects with larger standard errors), and points cor-
sults. This means that if an intervention seems to be responding to studies with larger sample size are scat-
more beneficial under the random-effects model than tered in a narrow range of values at the top of the
under the fixed-effect model, the researchers should funnel (because they yield effects with smaller standard
investigate further whether they should attribute this errors). Instead of standard errors, we could have used
difference to the small-study effects alone (the interven- the sample size of the studies or the variance of the effect
tion was more effective in smaller studies) or to other sizes.4 However, only the standard errors can spread out
study characteristics.2 the points on the bottom of the funnel where the smaller
The forest plot in Figure 1 displays the meta-analysis studies are found and create a funnel-like pattern.4
results on the effectiveness of fluoride gel against pla- To determine whether there is publication bias or
cebo for preventing dental caries in children and adoles- small-study effects, we need to understand how the
cents under the random-effects and fixed-effect points are distributed. The symmetrical distribution of
models.3 The point estimates differ very slightly, and the points about the summary effect size is an indication
the confidence intervals overlap perfectly. Note how of the absence of possible small-study effect or publica-
a
tion bias. However, any asymmetrical distribution of the
Midwifery Research and Education Unit, Hanover Medical School, Hanover,
Germany. points may support the presence of possible small-study
b
Department of Orthodontics and Dentofacial Orthopedics, University of Bern, effect or publication bias.2,4 The typical pattern in the
Bern, Switzerland. presence of small-study effects is a prominent asymme-
Address correspondence to: Nikolaos Pandis, Department of Orthodontics and
Dentofacial Orthopedics, University of Bern, Freiburgstrasse 7, CH-3010 Bern, try at the bottom that progressively disappears as we
Switzerland; e-mail, npandis@yahoo.com. move up to larger studies.4
Am J Orthod Dentofacial Orthop 2021;159:248-51 Figure 2 illustrates the funnel plot of our example. The
0889-5406/$36.00
Ó 2020. effect sizes have been estimated using the fixed-effect
https://doi.org/10.1016/j.ajodo.2020.11.005 model. The black line displays the summary effect size,
248
Statistics and research design 249

Fig 1. A forest plot that illustrates the meta-analysis results under the random-effects model (overall)
and the fixed-effect model (overall fixed-effect estimate).

and the red dotted line refers to no effect. The diagonal


lines represent the pseudo 95% confidence limits around
the summary effect size for each standard error on the ver-
tical axis. In the absence of heterogeneity, 95% of the
studies should be scattered within the funnel as defined
by these diagonal lines. The asymmetry of the funnel
plot is evident; toward the bottom of the plot, there is
only 1 small study, whereas the majority of the studies
are scattered above the middle of the funnel plot. Four
studies are outside the pseudo 95% confidence limits.
The absence of smaller studies (equivalently, only 1 small
study) at the bottom of the plot and studies with small ef-
fects (on the right side of the plot) are strong indications of
possible publication bias in the meta-analysis results. In
Fig 2. Funnel plot on the effectiveness of fluoride gel this case, we suspect that the summary effect may be
against placebo for preventing dental caries in children biased. However, the absence of smaller studies might
and adolescents.
also indicate that the effectiveness of fluoride gel was

American Journal of Orthodontics and Dentofacial Orthopedics February 2021  Vol 159  Issue 2
250 Statistics and research design

likely to be investigated mainly in moderate and large present in the contours of statistical significance than in
studies. Therefore, publication bias cannot be perceived areas of statistical nonsignificance, the funnel plot will
as the only cause of funnel asymmetry.2 be asymmetric.6 This asymmetry is indicative of the
Although funnel plot asymmetry has been associated inherent association between effect size and level of sig-
mostly with publication bias and small-study effects, nificance.6 Studies that tend to favor the experimental
they are not the sole reasons for an asymmetric funnel than the control intervention are more likely to provide
plot.2 Was the literature search intensive and compre- statistically significant results, and hence, to be present
hensive? Is there any difference in the quality of the in the contours of statistical significance. By contrast, if
studies? For instance, it has been shown that small more studies are absent in the contours of statistical sig-
studies with poor design and conduct tend to overesti- nificance than in areas of statistical nonsignificance,
mate the effect sizes; hence, this lower methodological then the risk of publication bias might be low. Figure 3 de-
quality of smaller studies might contribute to the asym- picts the contour-enhanced funnel plot of our example.
metry at the bottom of the funnel plot.2 Is there a plau- The black line displays the summary effect size, and the
sible reason to explain the larger effect sizes in smaller red dotted line refers to no effect. The gray-scaled con-
trials? For instance, the choice of the measure of effect tours correspond to different levels of significance, as indi-
can also result in funnel plot asymmetry,2 It has been cated in the accompanying box.
shown that risk difference provides more heterogeneous Studies are missing in all levels of significance both
effect sizes than risk ratio or odds ratio because they on the right side and at the bottom of the plot, creating
tend to underestimate the effect sizes in studies with pronounced funnel plot asymmetry. The fact that mostly
low event rates5; hence, they can result in funnel plot moderate and large studies with statistically significant
asymmetry. Therefore, various reasons related to study larger effect sizes are present indicates that publication
characteristics and/or the analysis of the study results bias may be the cause of funnel asymmetry; hence,
should be considered before we attribute the asymmetry further investigation is needed to detect possible associ-
entirely to publication bias.2,6 Meta-regression and sub- ations between effect size and study characteristics.
group analysis are the tools to investigate possible asso-
ciations between study characteristics and effect sizes. EGGER TEST FOR FUNNEL PLOT ASYMMETRY
Publication bias has been equated mainly with the sup- Because the interpretation of the funnel plot is sub-
pression of studies with statistically nonsignificant results. jective, statistical tests that assess the relationship be-
The funnel plot in Figure 2 fails to display the level of sta- tween sample size and effect size have been
tistical significance of the effect sizes to assess whether the developed. The test for small-study effects proposed
level of significance is likely to explain the asymmetry.6 A by Egger et al2 is the most widely used approach
contour-enhanced funnel plot is an extension of the con- (Fig 4). A regression line of the effect sizes against their
ventional funnel plot, which incorporates contours of the standard errors (using the inverse variance weighting
level of significance of the effect sizes.6 If more studies are scheme) is drawn on the funnel plot.2 If there is no asso-
ciation between effect size and standard error, the
regression line will be parallel to the x-axis.2 To examine
whether there is evidence for statistically significant

Fig 3. Contour-enhanced funnel plot on the effectiveness


of fluoride gel against placebo for preventing dental caries
in children and adolescents. Contours indicate the level of
significance of the effect sizes. Fig 4. Egger test for asymmetry.

February 2021  Vol 159  Issue 2 American Journal of Orthodontics and Dentofacial Orthopedics
Statistics and research design 251

asymmetry, we can perform a statistical test, known as Copas probabilistic model.9,10 A brief description of
Egger's test. Specifically, we test whether the intercept these approaches is provided below.
is equal to 0 which implies that the regression line A method known as “trim and fill” is implemented to
runs through the origin of the plot, and hence, the fun- understand the impact of publication bias.9 This method
nel plot is symmetrical. If the P value is below 0.10, we is an iterative procedure that identifies and corrects the
may conclude that there is evidence asymmetry because asymmetry in the funnel plot. It removes the small studies
of small-study effects.2 with the most extreme results (trim) and recalculates the
However, this test suffers the caveat of any test of summary effect size at each iteration until the funnel
significance: it yields results of low power.1 Therefore, plot becomes symmetric.9 Then, the removed studies are
the rejection of the null hypothesis should not be consid- added back into the analysis, and a “mirror value” is
ered as definite evidence of asymmetry. In addition, tests computed for each one (fill). The output of this method
for funnel plot asymmetry should be performed when is an “adjusted” funnel plot that depicts both the observed
there are at least 10 studies in the meta-analysis; other- and the imputed studies so that the researcher can see how
wise, the power will be too low to detect any true rela- much the summary effect size changes when the imputed
tionship.1 These tests should not be implemented studies are included (trivially, modestly, or substantially).9
when the studies have similar sample sizes.1 Another approach to assess the impact of publication
Please note that small-study effects are not the only bias is to use probabilistic models. Copas10 examined the
source of asymmetry in funnel plots; language bias, cita- relationship between the standard error of a study and the
tion bias, true heterogeneity, poor methodological probability that this study will be included in a meta-
design of small studies, poor choice of effect measures, analysis. A sensitivity analysis is performed to estimate
and chance are also potential sources of asymmetry.2 the summary effect sizes under a range of assumptions
about the probability of publication. Then the researcher
JUDGMENT ON THE ROBUSTNESS OF THE META- can explore how the estimated summary effect sizes and
ANALYSIS EFFECT SIZE their confidence intervals vary across the different sce-
In the presence of publication bias, the meta-analysis narios assumed in the sensitivity analysis and determine
results will not reflect reality. If we could retrieve all the the impact of publication bias in the results.
missing studies, then the direction or the statistical sig-
nificance of the results might be different.1 Approaches REFERENCES
to investigate the robustness of the summary effect size 1. Boutron I, Page MJ, Higgins JPT, Altman DG, Lundh A,
have been proposed by Rosenthal and Orwin.7,8 These Hrobjartsson A. Chapter 7. Considering bias and conflicts of inter-
approaches are presented very briefly below. est among the included studies. In: Higgins JPT, Thomas J,
According to Rosenthal, we could estimate the num- Chandler J, Cumpston M, Li T, Page MJ, et al, editors. Cochrane
Handbook for Systematic Reviews of Interventions. Chichester,
ber of (additional ‘negative’) studies needed to be
United Kingdom: Cochrane; 2019.
retrieved and included in the meta-analysis to increase 2. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-
the P value for the meta-analysis above 0.05.7 If we analysis detected by a simple, graphical test. BMJ 1997;315:629-34.
need only a small number of studies, we should be con- 3. Marinho VCC, Higgins JP, Logan S, Sheiham A. Fluoride gels for
cerned about the robustness of the summary effect size. preventing dental caries in children and adolescents. Cochrane
Database Syst Rev 2002;(2):CD002280.
This approach is known as Fail-safe N.
4. Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Chapter 30.
Orwin's approach is an extension of Rosenthal's strat- Publication bias. In: Introduction to Meta-Analysis. Chichester,
egy: researchers can determine the number of missing United Kingdom: John Wiley & Sons; 2009.
studies needed to bring the mean effect size below a spe- 5. Deeks JJ. Issues in the selection of a summary statistic for meta-
cific value other than 0 (for continuous measures) or 1 analysis of clinical trials with binary outcomes. Stat Med 2002;
21:1575-600.
(for ratio measures).8 This specific value would be an ef-
6. Peters JL, Sutton AJ, Jones DR, Abrams KR, Rushton L. Contour-
fect size that indicates a minimum clinically important enhanced meta-analysis funnel plots help distinguish publication
difference between the compared interventions. The re- bias from other causes of asymmetry. J Clin Epidemiol 2008;61:991-6.
searchers could also specify the mean effect size in the 7. Rosenthal R. The file drawer problem and tolerance for null results.
missing studies to be any value other than 0 or 1. Psychol Bull 1979;86:638-41.
8. Orwin RG. A fail-safe N for effect size in meta-analysis. J Educ Be-
hav Stat 1983;8:157-9.
JUDGMENT ON THE IMPACT OF PUBLICATION BIAS 9. Duval S, Tweedie R. Trim and fill: a simple funnel-plot-based
Two popular approaches to judge the impact of pub- method of testing and adjusting for publication bias in meta-anal-
ysis. Biometrics 2000;56:455-63.
lication bias on the conclusions drawn from the meta-
10. Copas J. What works?: selectivity models and meta-analysis. J R
analysis results are the trim-and-fill approach and the Stat Soc A 1999;162:95-109.

American Journal of Orthodontics and Dentofacial Orthopedics February 2021  Vol 159  Issue 2

You might also like