[Downloaded free from http://www.j-pcs.org on Saturday, February 22, 2020, IP: 1.39.135.
129]
Curriculum in Cardiology
Statistics
How to Read a Forest Plot
Ushmita Seth
Technology Consultant, B Tech ( Delhi Technological University), Delhi, India
Abstract
When the data‑based practice began to accumulate, forest plots were introduced to realize the collective power of the statistical data. It is a
graphical representation of a meta‑analysis, also known as blobbogram. It allows you to view and analyze the resulting sample individual
statistics from multiple similar studies all in one place, along with summary statistics at the bottom. The plot includes the point value of
the sample statistic as well as its confidence interval (usually taken as 95%).
Keywords: Forest plot, point estimate, statistics
A forest plot is a graphical display of one common statistical The pooling of diverse statistical analysis is done by two
conclusion from a number of studies directing the same methods either using fixed‑effects model or random‑effects
problem. This tackles the complexities of collective inferences model.[3] It has been recommended to use the random‑effects
of various experiments which lead to a powerful conclusion. pooling model in clinical psychology and the health sciences.[4]
The fixed‑effects model assumes that all studies are conducted
In 1990, oncologist Richard Peto joked that the plot was named on a single homogeneous population. While pooling the effect
after fellow breast cancer researcher Pat Forrest, resulting in the sizes, a weighted average of a sample statistic is conducted
frequent misnaming of the plot as Forest plot. However, it was with the study with smaller variance (i.e., greater precision)
named as the graph had a resemblance to an image of a forest given a larger weight.
when placed at a right angle [Figure 1]. As the plot consists
of lines and large dots, somewhere along each line, the line However, in practice, all studies can almost never be from
represents a tree and the dot corresponds to the leaf cover.[1] the same population, and therefore, alternatively, we can do it
using the random‑effects model. Here, we assume that studies
Let us understand the different branches of a forest plot with are conducted not only on one single population but also on
the given example [Table 1]. a “diverse” population. We, therefore, assume that there is
Here is a common representation of the raw data[2] for the plot. not only one true effect size but also a distribution of true
The first column signifies the name of the study. The second and effect sizes. We, therefore, want to estimate the mean of this
third columns describe the experimental results for treatment distribution of true effect sizes.
and control groups, respectively. “n” stands for the number θk = θF+ϵk + ζk
of patients who had the outcome, and “N” stands for the total
θk = Observed effect size of an individual study k
number of people in the group.
θF = True effect size of the population
The third column generally indicates the point estimate of the
common statistic that is being used to compare all the studies. Address for correspondence: Ms. Ushmita Seth,
It could be a relative statistic, such as odds ratio (OR) or E‑mail: [email protected]
relative risk (RR), or it could be an absolute statistic such as
standardized mean difference or absolute risk reduction. The Date of Submission: 21‑Jun‑2019 Date of Acceptance: 28-Jul-2019
fourth and fifth columns represent the upper and lower bounds Date of Revision: 10-Jul-2019 Date of Web Publication: 19-Aug-2019
of the confidence interval (CI), respectively.
This is an open access journal, and articles are distributed under the terms of the Creative
Access this article online Commons Attribution‑NonCommercial‑ShareAlike 4.0 License, which allows others to
remix, tweak, and build upon the work non‑commercially, as long as appropriate credit
Quick Response Code:
is given and the new creations are licensed under the identical terms.
Website:
www.j‑pcs.org
For reprints contact:
[email protected] DOI:
10.4103/jpcs.jpcs_39_19 How to cite this article: Seth U. How to read a forest plot. J Pract Cardiovasc
Sci 2019;5:108-10.
108 © 2019 Journal of the Practice of Cardiovascular Sciences | Published by Wolters Kluwer - Medknow
[Downloaded free from http://www.j-pcs.org on Saturday, February 22, 2020, IP: 1.39.135.129]
Seth: Forest plot
ϵk = Sampling error even be the true value. Therefore, the study is not statistically
significant. Forest plot indicates the estimated effects of CIs for
ζk = Second type of error as even the true effect size θF is
individual study and also overall estimated effects of CIs.
also a part of distribution of true effect sizes (of the universe
of populations) The diamond at the bottom represents the summary statistic and
CI based on a meta‑analysis. The center of the diamond (or if you
To take ζk into account, we have to estimate the variance of
draw a vertical line joining its vertical points) represents the point
the distribution of true effect sizes, which is denoted by τ2, or
estimate. The horizontal points represent the CIs. As the diamond
tau2. There are several estimators for τ2.
is a culmination of all the individual studies, the CI would be
As in fixed‑effects model, we require a weight to be assigned the smallest (CIs are inversely proportional to sample size, as
to each study which would decide its influence on the overall larger sample size means smaller standard error and vice versa).
meta‑analysis. The choice of estimator defines the final
The final point about analyzing a forest plot is its “heterogeneity.”
calculation of the variance and, therefore, leads to different
Heterogeneity arises due to the bias creeping into the final estimate
pooled sized estimates and CIs. An article by Veroniki et al.[5]
as the individual studies have been conducted using different
provides a summary of various estimators and their biases.
methods across different populations. Therefore, an additional
Let us now draw the forest plot corresponding to the above commonly used metric called “I2” or I‑squared[6] is calculated at
data [Figure 1]. the end of the plot. If I2 is <50%, then the individual studies fall
First, we look at the two axes. The X‑axis is the scale for the within the acceptable range of inconsistency. If it is >50%, then
statistics being displayed (OR in our case). The vertical line they are too inconsistent to be used together for the meta‑analysis.
is not a Y‑axis as such; it is the line of “null effect” for the
statistic which has been used in our case – the value of the point Conclusion
statistic which signifies no difference between treatment and A forest plot is a graphical display of results from a number
control groups. It would be placed at 1 for a relative statistic of studies addressing the same question. It is called a forest
and at 0 for an absolute statistic. plot [Figure 2] because it represents a forest of lines. It
Next, the results of each study are placed one below the other was developed as a means of graphically representing a
on the plot. For each study, the location of square with respect meta‑analysis. They are commonly presented with two
to X‑axis marks its point estimate, the size of the square marks columns. The left‑hand column lists the names of the studies.
the sample size, and the length of the horizontal line on which The right‑hand column is a plot of the measure of effect (e.g.,
the square lies represents the CI for the point estimate. If at any RR) for each of these studies, represented by a square,
point, the horizontal line crosses the line of null effect, it basically
means that the point of null effect lies within your CI and could
Figure 1: The plot is drawn using RStudio Version 1.1.463. Figure 2: Forest plot showing benefit.
Table 1: Let us understand the different branches of a forest plot with the given example
Study Treatment (n/N) Control (n/N) Point estimate Weight 95% CI
IDs (e.g., OR) (%)
Lower Upper
Study 1 1/131 2/133 0.5 17.8 0.05 5.49
Study 2 7/279 9/290 0.84 77.7 0.36 1.93
Study 3 3/102 1/101 3.00 4.5 0.12 72.77
Total 512 (sum of total members of all groups) 542 (sum of total members of all groups) 0.87 100 0.41 1.87
Data were obtained from students4bestevidence.com. OR: Odds ratio, CI: Confidence interval
Journal of the Practice of Cardiovascular Sciences ¦ Volume 5 ¦ Issue 2 ¦ May-August 2019 109
[Downloaded free from http://www.j-pcs.org on Saturday, February 22, 2020, IP: 1.39.135.129]
Seth: Forest plot
incorporating CIs represented by horizontal lines. The overall References
measure of effect is represented as a dashed vertical line. This 1. Lewis S, Clarke M. Forest plots: Trying to see the wood and the trees.
is plotted as a diamond, the lateral points of which indicate BMJ 2001;322:1479‑80.
2. Available from: https://www.students4bestevidence.net/wp‑content/
CIs for this estimate. A vertical line representing no effect is uploads/2016/06/How‑to‑read‑a‑forest‑plot‑2.jpg. [Last accessed on
also plotted, and if the points of the diamond overlap the line 2019 Jul 08].
3. Borenstein M, Hedges LV, Higgins JP, Rothstein HR. Introduction to
of no effect, the overall result cannot be said to differ from no
Meta‑Analysis. United Kingdom: John Wiley & Sons; 2011.
effect at the given level of confidence. 4. Cuijpers P. Meta‑Analyses in Mental Health Research. A Practical
Guide. 2016. Available from: https://indd.adobe.com/view/5fc8f9a0-
Financial support and sponsorship bf1e-49d3-bf5f-a40bfe5409e0. [Last accessed on 2019 Jul 08].
5. Veroniki AA, Jackson D, Viechtbauer W, Bender R, Bowden J, Knapp G.
Nil.
Methods to estimate the between‑study variance and its uncertainty in
meta‑analysis. Res Synth Methods 2016;7:55‑79.
Conflicts of interest 6. Higgins JP, Thompson SG. Quantifying heterogeneity in a meta‑analysis.
There are no conflicts of interest. Stat Med 2002;21:1539‑58.
110 Journal of the Practice of Cardiovascular Sciences ¦ Volume 5 ¦ Issue 2 ¦ May-August 2019
