EBSJ Special Section: Science-in-Spine

Global Spine Journal

2021, Vol. 11(4) 614-616
Seeing the Forest by Looking at the Trees: ª The Author(s) 2021
Article reuse guidelines:

How to Interpret a Meta-Analysis Forest Plot sagepub.com/journals-permissions

DOI: 10.1177/21925682211003889
journals.sagepub.com/home/gsj

Joseph R. Dettori, PhD1, Daniel C. Norvell, PhD1,

and Jens R. Chapman, MD2

A forest plot is a useful graphical display of findings from a (effect size; e.g. risk ratio, odds ratio, or mean difference), and
meta-analysis. It provides essential information to inform our its size (area) is proportionate to the weight of the study. Not all
interpretation of the results. Typically, a forest plot contains 6 studies contribute equally to the pooled results. In general,
basic “columns”, though additional columns can be added to studies that have a larger N provide more information and are
provide more information. The 6 basic columns include details therefore allotted greater weight. The design draws our eyes
relating to the following: toward the studies that are given more weight. This is seen
readily in the study by Gornet et al in Figure 1 which has a
1. included studies (and subgroups if analyzed) sample size of 405 compared to 205 and 53 in Blumenthal and
2. intervention group Geisler, respectively.
3. control group Remember, the point estimate is the best guess of the true
4. weight effect in the population. The width of the study lines extending
5. outcome effect measure in numeric format
through the boxes shows their confidence intervals. The confi-
6. outcome effect measure in graphical presentation
dence interval represents the chance that the true effect in the
Let’s look at 2 examples from a study comparing total disc population will lie within the range.
replacement with anterior lumbar interbody fusion to discover The diamond below the studies represents the overall pooled
the usefulness of forest plots.1 effect from the included studies. The width of the diamond
shows the confidence interval for the overall effect.
Each forest plot contains a vertical line, the line of ‘no
Begin at the End to Get Your Bearings effect’, which corresponds to the value 1 for binary outcomes
We suggest first looking at the type of outcome found in such as the risk ratio or odds ratio and 0 in the case of contin-
columns 5 and 6 as it influences the contents in other columns uous outcomes. When the 95% CI from a single study or the
(Figures 1 and 2). If the outcome is binary, the number of pooled estimate crosses the line of no effect, the difference in
events (numerator) and the total (denominator representing outcome between intervention and comparator is not statisti-
the population size) will be presented in columns 2 and 3, cally significant. Otherwise, statistical significance exists. In
Figure 1. With binary data, the ratio between risks (risk ratio) Figure 1, the pooled point estimate and the 95% CI lies entirely
or odds (odds ratio) are calculated and presented numerically to the right of the line of no effect. This tells us that there is a
in column 5 and graphically in column 6 along with its 95% statistical difference in the outcome between groups. In this
confidence interval (95% CI). However, if the outcome is figure, the results of satisfaction favor the ALIF group. It is
continuous, the difference in the means (mean difference) confirmed by the test for overall effect located at the bottom
between the intervention and control groups is displayed in left of Figure 1, P ¼ .001. On the other hand, the diamond in
columns 5 and 6 with its 95% CI, Figure 2. And, the individual
study means, standard deviations (sd) and sample sizes are
1
given in columns 2 and 3. Spectrum Research, Inc, Steilacoom, WA, USA
2
Swedish Neuroscience Institute, Seattle, WA, USA

Understand the Graphical Display Corresponding Author:

Joseph R. Dettori, Spectrum Research, Inc, PO Box 88998, Steilacoom,
Each line in the graphical display represents a study. The mid- WA 98405, USA.
point of the box symbolizes the point estimate of the effect Email: [email protected]

Creative Commons Non Commercial No Derivs CC BY-NC-ND: This article is distributed under the terms of the Creative Commons Attribution-Non
Commercial-NoDerivs 4.0 License (https://creativecommons.org/licenses/by-nc-nd/4.0/) which permits non-commercial use, reproduction and distribution of the
work as published without adaptation or alteration, without further permission provided the original work is attributed as specified on the SAGE and Open Access
pages (https://us.sagepub.com/en-us/nam/open-access-at-sage).
Dettori et al 615

Figure 1. Proportion of patients satisfied with total disc replacement (TDR) versus anterior lumbar interbody fusion (ALIF) (Mu et al, 20 181).

Figure 2. Blood loss comparing TDR vs ALIF (Mu et al, 20 181).

Figure 2 crosses the line of no effect suggesting no statistically This statistical heterogeneity, often referred to simply as het-
significant difference. This is verified by the test of overall erogeneity, can be evaluated in 3 ways:
effect, P ¼ .15.
1. By gauging the overlap of the included studies’ point
estimates and their 95% confidence intervals.
Understand Heterogeneity 2. By looking at the P-value of the Chi.2
A forest plot provides information about the heterogeneity 3. By assessing the I2 test, which quantifies the magnitude
among studies. Since several primary studies are brought of the heterogeneity.
together to provide one estimate (represented by the diamond
in the forest plot), variability among them is inevitable. Clinical Compare Figures 1 and 2 for heterogeneity. The overlap of
heterogeneity (variability in participants, treatments and out- point estimates and confidence intervals in Figure 1 tend to be
comes) and methodological heterogeneity (variability in study more consistent compared with Figure 2. This is corroborated
design and risk of bias) can be reflected in statistical hetero- by the Chi2 test of heterogeneity that tests the hypothesis of no
geneity (variability in the treatment effects being evaluated). heterogeneity. P < .001 in Figure 2 rejects the hypothesis of no
616 Global Spine Journal 11(4)

heterogeneity whereas P ¼ .20 from Figure 1 does not reject outcomes (e.g., risk ratio or odds ratio) and 0 in the case
the hypothesis of no heterogeneity. The magnitude of hetero- of continuous outcomes.
geneity is estimated by the I2 and its interpretation is roughly as  When the 95% CI from a single study or the pooled
follows2: estimate crosses the line of no effect, the difference
between intervention and comparator is not statistically
0% to 40% Might not be important
significant. Otherwise, statistical significance exists.
30% to 60% May represent moderate heterogeneity  Heterogeneity among studies is inevitable, and its mag-
50% to 90% May represent substantial heterogeneity nitude is estimated by the I2 statistic.
75% to 100% Considerable heterogeneity

Declaration of Conflicting Interests

2
The I in Figure 1 is 38% suggesting any heterogeneity The author(s) declared no potential conflicts of interest with respect to
the research, authorship, and/or publication of this article.
might not be important, whereas the 87% in Figure 2 suggests
substantial heterogeneity.
Funding
The author(s) received no financial support for the research, author-
Summary ship, and/or publication of this article.
 Forest plots are useful graphical displays summarizing
results from a meta-analysis. References
 When interpreting a forest plot, first identify the type of 1. Mu X, Wei J, A J, Li Z, Ou Y. The short-term efficacy and safety of
outcome used (e.g., binary or continuous). artificial total disc replacement for selected patients with lumbar
 Each study included in a meta-analysis is represented by degenerative disc disease compared with anterior lumbar interbody
a box (point estimate) and a horizontal line through the fusion: a systematic review and meta-analysis. PLoS One. 2018;
box (95% confidence interval). The size of the box rep- 13(12):e0209660.
resents the study weight; the larger the box, the more 2. Deeks JJ, Higgins JPT, Altman DG, editors. Analysing data and
information the study provides and the greater the undertaking meta-analyses. In: Higgins JPT, Thomas J, Chandler J,
weight. The diamond below the studies represents the Cumpston M, Li T, Page MJ, Welch VA, editors. Cochrane Hand-
overall pooled effect from the included studies. book for Systematic Reviews of Interventions version 61 (updated
 Each forest plot contains a vertical line, the line of ‘no September 2020). Cochrane; 2020. Chapter 10. www.training.
effect’, which corresponds to the value 1 for binary cochrane.org/handbook