ACADEMIC WRITING GUIDE
How to Read and Report a Forest Plot in a Dissertation (With Worked Examples)
A forest plot can look decisive at first glance: several squares fall to one side of a vertical line, and a diamond at the bottom appears to favour an intervention. Yet a dissertation can describe that picture accurately and still draw the wrong conclusion. The plot shows estimated effects and uncertainty, not whether the underlying studies are unbiased, sufficiently comparable or applicable to a particular population.
This guide explains how to read, calculate, reproduce and report a forest plot in a systematic review dissertation. Instead of merely listing what its symbols mean, we work from a fictional three-study dataset to individual risk ratios, confidence intervals, inverse-variance weights and a pooled estimate. We then turn those results into defensible dissertation writing and show the checks that should precede any conclusion.
Teaching example and scope: All three studies and their outcome counts are entirely invented. The worked analysis uses a common-effect inverse-variance model on the log risk-ratio scale, with approximate normal-theory 95% confidence intervals. It is a demonstration of plot construction, not a recommendation to pool three real studies automatically, an analysis of actual intervention effectiveness or a substitute for your review protocol. The suitability of the model, precision method and underlying studies must be evaluated in a real dissertation. The mathematical approach follows the Cochrane Handbook's meta-analysis guidance and its effect-measure chapter.
What is a forest plot in a dissertation?
A forest plot is a display of effect estimates and their uncertainty from individual studies, often alongside a combined result from a meta-analysis. Each study normally occupies a row. The horizontal location of its marker represents the effect estimate; a line through the marker represents an interval such as a 95% confidence interval (CI). If a synthesis is calculated, a diamond commonly represents the pooled estimate and its interval. The visual convention is useful, but always read the figure legend: not every forest plot has the same model, weighting method, axis orientation or reported interval.
A forest plot is not the same as a funnel plot. The forest plot presents each study's estimated result and often a summary; a funnel plot investigates the relationship between effect estimates and study precision to help assess small-study effects. Neither figure proves the absence of publication bias.
Cochrane and PRISMA 2020 reporting guidance emphasize explaining which studies contributed to a synthesis, how estimates were produced and what limitations affect interpretation. A polished image cannot substitute for those decisions.
How to read the six core elements
| Element | Typical appearance | What it means | Common misreading |
|---|---|---|---|
| Study label | Name and year on the left | Identifies the source of an effect estimate | Assuming separate reports are separate studies |
| Point estimate | Square or dot | Estimated effect for that study | Treating the estimate as the true effect |
| Study interval | Horizontal line | Uncertainty around the estimate at the stated level | Calling a wide interval “no effect” |
| Weight | Square area or numerical percentage | Influence assigned under the specified synthesis model | Assuming weight equals study quality |
| Line of no effect | Vertical line at RR/OR = 1 or MD/SMD = 0 | Value corresponding to no difference on that scale | Using zero for every effect measure |
| Summary diamond | Centre and width of diamond | Combined estimate and its interval, if pooling was performed | Treating a pooled result as proof of causation |
1. Identify the effect measure before reading left and right
A risk ratio (RR) compares event risks. RR = 1 represents equal risk in the two groups. An RR below 1 means the event is less frequent in the group used in the numerator; whether that is favourable depends on whether the outcome is desirable or harmful. An RR above 1 means the numerator group's event risk is greater.
For an odds ratio (OR) or hazard ratio (HR), the conventional no-effect value is also 1, but those measures have different definitions and should not be called risk ratios. For a mean difference (MD), the no-effect value is typically 0, and the units are those of the original outcome. Standardized mean differences are also centred on 0 but express an effect in standard-deviation units. Check the model and units, not just the sign.
Many ratio-scale forest plots display a logarithmic horizontal axis, making values such as 0.5 and 2 equally far from 1. The displayed numbers may be back-transformed RRs even if calculations use log RRs. This is standard, not an inconsistency.
2. Interpret the study marker and interval together
The marker's position is the study estimate; the accompanying line is a measure of its precision. A narrow interval means greater precision under the assumptions used; it does not establish freedom from risk of bias. A wide interval can include substantial benefit and harm. If a 95% CI crosses the no-effect line, the data remain compatible with effects on both sides of that line at that confidence level; it does not show that the two interventions are equivalent.
A plotted interval should be reproduced from the correct effect measure, sampling variance and analysis method. Some forest plots instead show prediction intervals, Bayesian credible intervals or adjusted estimates. Explain which is displayed.
3. Check study weights and the pooled diamond
In a conventional inverse-variance common-effect analysis, more precisely estimated study effects receive larger weights. Other models assign different weights. A large square therefore does not indicate a study with low bias, an especially important outcome or greater applicability. It indicates greater statistical weight under that model.
The diamond's centre typically marks the pooled estimate and its horizontal span the corresponding CI. The pooled result is meaningful only if the review question, included studies, extraction, effect direction and synthesis method justify combining them. For the difference between a pooled average and the expected range of effects in a new setting, see our heterogeneity and prediction-interval guide.
Worked example: turn three fictional studies into a forest plot
Imagine a review asks whether a fictional student-support programme reduces course withdrawal compared with usual support. Withdrawal is the undesirable event. Each invented study randomizes 100 students to the programme and 100 to comparison. For teaching simplicity, assume independent groups, a single withdrawal outcome, complete follow-up and no zero-event cells.
| Study | Programme withdrawals / total | Comparison withdrawals / total | Programme risk | Comparison risk |
|---|---|---|---|---|
| Study A | 20 / 100 | 30 / 100 | 0.20 | 0.30 |
| Study B | 12 / 100 | 20 / 100 | 0.12 | 0.20 |
| Study C | 30 / 100 | 40 / 100 | 0.30 | 0.40 |
These are synthetic counts, not results retrieved from papers. In a real review, verify that the event definition, populations, comparison and follow-up time are sufficiently compatible, distinguish multiple reports of the same study, and check outcome-level risk of bias before pooling. Acadelyra's risk-of-bias guide covers that separate judgment.
Step 1: Calculate the study risk ratios
Risk is events divided by group size. A study RR is programme risk divided by comparison risk.
For Study A, programme risk is 20 ÷ 100 = 0.20 and comparison risk is 30 ÷ 100 = 0.30. Its RR is therefore 0.20 ÷ 0.30 = 0.667 (three-decimal rounding). The estimated withdrawal risk in the programme group is about one-third lower relative to comparison for this fictional study. That is an interpretation of the estimate, not a proven causal conclusion.
For Study B, the RR is 0.12 ÷ 0.20 = 0.600. For Study C, the RR is 0.30 ÷ 0.40 = 0.750. All three estimated RRs are below 1, but all require confidence intervals before we can discuss precision.
Step 2: Calculate the uncertainty for each study
With event counts a and c and group totals n₁ and n₀, a conventional large-sample standard error of the natural log RR is:
SE[ln(RR)] = sqrt(1/a − 1/n₁ + 1/c − 1/n₀)
The approximate 95% interval on the log scale is ln(RR) ± 1.96 × SE; exponentiate both endpoints to return to the RR scale. This example contains no zero cells. When data are sparse or zero events occur, analysts must select and report a suitable method rather than apply this formula mechanically.
For Study A, SE = sqrt(1/20 − 1/100 + 1/30 − 1/100) = 0.2517 (rounded). Its log RR is ln(20/30) ≈ −0.4055. The approximate RR interval is 0.407 to 1.092 after exponentiating the log-scale endpoints. It crosses 1, so the individual study remains compatible with no difference, as well as meaningful relative reductions in withdrawal.
For Study B, the approximate interval is 0.310 to 1.161. For Study C, it is 0.511 to 1.101. Each of these individual 95% intervals includes 1. A forest plot should make that uncertainty visible rather than reporting only the direction of the point estimates.
Step 3: Understand inverse-variance weights
To illustrate a common-effect inverse-variance synthesis, transform each RR to its natural logarithm, calculate the variance as the squared log-scale standard error, and use weight = 1 / variance. The study with the smaller variance receives more weight.
For Study A the unnormalized weight is 15.789; for Study B it is 8.824; and for Study C it is 26.087. The unnormalized weights sum to 50.700. When divided by that sum, their plot weights are about 31.14%, 17.40% and 51.45%, respectively (displayed rounding may not sum to exactly 100%). Study C contributes the most precision under this model, not necessarily the most trustworthy evidence.
Step 4: Calculate and interpret the summary diamond
The pooled log RR is the sum of each study's log RR multiplied by its inverse-variance weight, divided by the sum of weights. For this dataset, the pooled log RR is approximately −0.3632. Exponentiating gives pooled RR = 0.695. The pooled log-scale standard error is 0.1404 (rounded), giving an approximate 95% CI of 0.528 to 0.916 on the risk-ratio scale.
Under the assumed common-effect model, the pooled estimate therefore corresponds to an estimated relative withdrawal reduction of approximately 30.5% ((1 − 0.695) × 100), not a 30.5-percentage-point absolute reduction. The pooled CI does not cross 1, whereas all three individual study intervals do. That is possible: combining independent information can increase precision. It does not mean the individual studies were “wrong,” nor does it remove concerns about bias, comparability or selective availability.
| Row | Risk ratio | Approximate 95% CI | Common-effect weight | Plot interpretation |
|---|---|---|---|---|
| Study A | 0.667 | 0.407–1.092 | 31.14% | Point left of 1; interval crosses 1 |
| Study B | 0.600 | 0.310–1.161 | 17.40% | Point left of 1; interval crosses 1 |
| Study C | 0.750 | 0.511–1.101 | 51.45% | Largest statistical weight; interval crosses 1 |
| Common-effect pooled | 0.695 | 0.528–0.916 | 100% combined | Diamond entirely below 1 |
How to draw the forest plot: Place Studies A–C in separate rows; mark each RR along a ratio-scale x-axis with a vertical no-effect line at 1; add its two CI endpoints as a horizontal line; size each study square according to the specified weights; place the summary diamond centred at 0.695 with tips at 0.528 and 0.916. Label the outcome as withdrawal, the numerator as programme, and explicitly state whether lower values favour the programme. Use a logarithmic x-axis so the ratio scale is not distorted. Do not interpret a graphical sketch as a substitute for plotted values checked against the extraction sheet.
Step 5: Reproduce the actual forest plot in R
For a reproducible plot, the following example uses the documented metafor R package to calculate log risk ratios and their variances from event counts, fit a common-effect model and display the forest plot. This is an optional teaching script; the Markdown article deliberately does not embed a fragile text-art plot that may misalign on mobile screens.
# Fictional teaching data. Requires the metafor package in R.
library(metafor)
example <- data.frame(
study = c("Study A", "Study B", "Study C"),
programme_events = c(20, 12, 30),
programme_nonevents = c(80, 88, 70),
comparison_events = c(30, 20, 40),
comparison_nonevents = c(70, 80, 60)
)
es <- escalc(
measure = "RR",
ai = programme_events, bi = programme_nonevents,
ci = comparison_events, di = comparison_nonevents,
data = example
)
model <- rma(yi, vi, data = es, method = "FE")
forest(
model, slab = example$study,
plim = c(NA, 1.5),
atransf = exp, refline = 0,
at = log(c(0.25, 0.5, 1, 2)),
xlab = "Risk ratio (programme / comparison)"
)
summary(model)Here, measure = "RR" produces log risk ratios and variances; atransf = exp displays axis labels and results on the risk-ratio scale, while refline = 0 sets the no-effect line on the underlying log scale. The common-effect method = "FE" was selected only to reproduce the teaching arithmetic. A different estimator, model, handling of zero cells or small-sample CI method may produce different estimates and intervals, and should be justified in a real review. Consult the metafor model documentation and keep the software version, complete data and analysis script with the dissertation's reproducibility materials.
Why a “statistically significant diamond” is not enough
There are several reasons this worked plot cannot support a claim that the fictional intervention is effective in real students.
Only three studies appear in the example. With few studies, heterogeneity estimates can be imprecise and random-effects uncertainty difficult to estimate robustly. The convenience of a common-effect calculation does not demonstrate that all study effects share one common value. In actual analyses, inspect clinical and methodological differences and justify whether pooling is appropriate in the first place.
Statistical weights are not quality scores. High precision cannot repair biased allocation, selective outcome reporting, inappropriate adjustment or missing outcomes. Apply a review-design-appropriate assessment to the outcome and discuss how those judgments affect the synthesis.
The plot may omit relevant evidence. Publication bias, unavailable results and selective inclusion can distort the body of evidence behind a visually persuasive summary. Use a risk-of-bias assessment alongside relevant missing-evidence methods rather than claiming that symmetrical study positions establish completeness.
The pooled RR is not an absolute reduction. An RR of 0.695 is a ratio. To express an absolute effect, specify a defensible baseline risk and calculate the expected difference; a single absolute claim without context could mislead. The uncertainty in baseline risk and effect must also be considered.
The confidence interval concerns the model's summary parameter, not effects in every new setting. A common-effect pooled CI is not a prediction interval. Where appropriate, random-effects modelling and a prediction interval help communicate variation across populations, but require adequate data and careful assumptions. For a deeper treatment, assess between-study variation and consider appropriate prediction intervals when sufficiently supported.
Certainty belongs to the body of evidence. A narrow or non-null pooled interval does not establish high certainty. If the review is an intervention-effect synthesis, an outcome-level GRADE assessment may incorporate risk of bias, inconsistency, indirectness, imprecision and publication bias. See our GRADE and Summary of Findings guide.
How to report a forest plot in dissertation Methods and Results
Methods: say precisely how the plot and synthesis were made
A useful Methods paragraph names the outcome, comparison, effect measure, handling of scales, eligibility for pooling, synthesis model, uncertainty method, software, treatment of missing statistics and planned examinations of heterogeneity. It should describe decisions you actually made, not present the teaching example as completed analysis.
Illustrative Methods paragraph:
For each eligible study, we extracted the numbers of participants experiencing the prespecified binary outcome and the total participants in each group. Risk ratios and 95% confidence intervals were calculated on the natural-log scale. For the illustrative synthesis, estimates were combined with a common-effect inverse-variance method; forest plots display individual study estimates, corresponding intervals and the pooled result. We checked consistency of outcome definitions, group orientation and follow-up timing before synthesis. Analyses were reproduced from recorded counts in R using the metafor package. The appropriateness of a common-effect assumption was considered separately from the numerical pooling calculation.
This paragraph is a model. A real dissertation must document its actual choices, including software version and any deviations from the protocol.
Results: report the numbers, then interpret them with limits
Readers should not have to infer all findings from a picture. Give the studies and participants in the synthesis, the pooled effect and interval, the chosen model and important qualifications. If heterogeneity or bias was assessed, report those results separately; do not invent I² values because a forest plot was produced.
Illustrative Results paragraph:
Three fictional studies, each comparing a student-support programme with usual support, contributed data on course withdrawal. Individual risk ratios ranged from 0.600 to 0.750, and each study's approximate 95% confidence interval included 1. In the prespecified teaching common-effect inverse-variance model, the pooled RR was 0.695 (95% CI 0.528–0.916). The summary interval excluded 1, although the small number of studies and the untested assumption of a common effect substantially limit the interpretation. These values illustrate statistical reporting and do not constitute evidence that any real programme reduces withdrawal.
A dissertation with real data should also describe the underlying comparability, risk-of-bias findings, non-contributing eligible studies and outcome-specific certainty judgments. Claims about impact must reflect evidence—not the formatting of the diamond.
Figure caption: make the forest plot stand alone
Figure: Fictional forest plot of course withdrawal risk ratios. Three invented studies are displayed as risk ratios with approximate 95% CIs. Square area represents inverse-variance weight under an illustrative common-effect model. The vertical no-effect line corresponds to RR = 1; lower risk ratios indicate fewer withdrawals in the programme group. The pooled diamond denotes RR 0.695 (95% CI 0.528–0.916). Figures and counts are teaching data, not empirical research results.
A strong figure legend should also note any transformations, subgroup divisions, unusual weighting, clipped intervals, omitted studies or departures from a conventional layout.
Eight forest-plot mistakes to avoid
| Mistake | Why it is a problem | Better approach |
|---|---|---|
| Calling a risk ratio an odds ratio | They quantify different comparisons | Label the correct measure and its scale |
| Saying a CI crossing 1 proves no effect | It may include meaningful effects in either direction | Describe compatible values and uncertainty |
| Calling the diamond a certainty grade | Statistical precision is not evidence certainty | Assess risk of bias and GRADE as appropriate |
| Reading a larger square as higher quality | Weight is model-dependent precision | Report appraisal separately |
| Pooling incompatible outcomes | A precise average can be uninterpretable | Group comparable outcomes or use structured synthesis |
| Using a linear axis for logarithmic ratio interpretation without clarity | Symmetry and visual distances can mislead | Label transformations and plotting scale |
| Reporting only favourable outcomes | Selective synthesis can distort conclusions | Account for planned outcomes and missing evidence |
| Calling a subgroup's different point estimates a confirmed interaction | Visual difference is not an interaction test | Use a justified formal comparison and cautious wording |
That last point matters when a forest plot contains subgroups. Two subgroup diamonds can look different without statistically reliable evidence that the intervention effect actually varies between groups. Our subgroup analysis and meta-regression guide explains formal interaction comparisons and their limitations.
When should you display a forest plot without a pooled diamond?
Sometimes displaying individual study estimates is helpful even when combining them would be misleading. Incompatible interventions, outcome definitions, follow-up intervals, designs or severe methodological concerns may make a single summary estimate inappropriate. A forest plot can still present comparable estimates and uncertainty without implying that a pooled effect has been justified. It must be clear in the Methods and figure legend why no diamond appears.
When numerical estimates cannot reasonably be put on a common scale, a structured evidence table or another form of transparent synthesis may be more appropriate. Acadelyra's narrative synthesis guide explains the alternatives. Avoid vote counting by statistical significance or treating the number of favourable studies as a pooled effect.
Frequently asked questions
Is a forest plot mandatory in every systematic review dissertation?
No. It is often valuable for displaying study estimates and meta-analytic results, but the appropriate display follows the review question, available data, synthesis method and university requirements. A review without pooled effects can use a forest plot without a summary diamond or an alternative display when better justified.
What does it mean if the pooled diamond crosses the no-effect line?
For a conventional 95% CI, the summary data and model remain compatible with the null value. The interval may still include clinically or educationally important effects. Report the range and substantive implications rather than simply stating “no difference.”
Can individual CIs cross 1 while the pooled CI does not?
Yes. Our three-study example illustrates that a combined estimate can be more precise than its contributors under a specified statistical model. That does not establish that pooling is methodologically justified or eliminate uncertainty about bias and heterogeneity.
Are RR and OR interchangeable on a forest plot?
No. Both have a no-effect value of 1 but describe different comparisons. An odds ratio can diverge substantially from a risk ratio when an outcome is common. Check the effect measure and source data before interpreting magnitude.
Does every forest plot need I² or a prediction interval?
Not automatically. Statistical heterogeneity and possible effect variation should be considered whenever studies are combined, but the presentation and methods depend on design, model, number of studies and available information. Avoid reporting I² or a prediction interval without an appropriate calculation and explanation.
How do I know whether a figure is reproducible?
Keep the extraction table, study identifiers, event counts or effect estimates, analysis scripts, software versions, transformations and plotted labels. A second person should be able to reconstruct both the numerical table and the graphical plot using those materials. If a graphical point appears inconsistent with the table, resolve the discrepancy before submission.
Final checklist before submitting the dissertation
Check the effect measure and event direction; confirm study identity and outcome comparability; reproduce every study estimate and CI; identify the weighting method; verify the summary model and diamond; label the null value correctly; disclose any omitted or ineligible data; assess bias and heterogeneity; distinguish confidence in the statistical estimate from certainty of evidence; report the numbers in text; and retain the extraction data and code necessary to regenerate the plot.
Final takeaway
A good forest plot is not an attractive diamond at the bottom of a dissertation page. It is a transparent link between the question, the included studies, the underlying data, the analysis and the interpretation. Read its axes and uncertainty before its apparent direction; understand what the model assumes before describing the pooled result; and report the limitations that the plot cannot display.
If you would like expert human help reviewing the clarity of your evidence tables, method descriptions, effect-size interpretation or dissertation writing, explore Acadelyra's Research Support and Dissertation & Thesis Support services. Our role is to support clear, verifiable academic work while the research decisions, data and submitted work remain your own.
Authoritative references and further reading
Cochrane. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 6: Choosing effect measures and computing estimates of effect. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-06
Cochrane. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 9: Summarizing study characteristics and preparing for synthesis. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-09
Cochrane. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 10: Analysing data and undertaking meta-analyses. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10
Cochrane. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 15: Interpreting results and drawing conclusions. https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-15
Page, M. J., et al. (2021). PRISMA 2020 explanation and elaboration: Updated guidance and exemplars for reporting systematic reviews. BMJ, 372, n160. https://pmc.ncbi.nlm.nih.gov/articles/PMC8005925/
Viechtbauer, W. metafor: Calculate effect sizes and outcome measures (
escalc) — package reference. https://wviechtb.github.io/metafor/reference/escalc.htmlViechtbauer, W. metafor:
rma.unimodel reference. https://rdrr.io/cran/metafor/man/rma.uni.html