Medical research methods · Forest plots
How to read a forest plot: effect sizes, 95% confidence intervals and pooled results
Turn a forest plot into an evidence-based results interpretation Requested outputs / delivery goal: An effect-measure checking table, an interpretation draft and a list of unresolved questions.
Identify what each row represents, then read the effect measure, null value, outcome direction and confidence interval before interpreting synthesis and heterogeneity. A plot displays estimates and uncertainty; it alone cannot establish causality, study quality or clinical importance.
This guide focuses on conventional meta-analysis forest plots in systematic reviews. Cohort models and subgroups can use similar graphics, but rows need not be independent studies and a pooled effect may not exist. One set of rules does not fit every forest plot.
Open the methods, results and legend together. The study-selection guide covers inclusion; this page explains reported results. It is not a statistical execution tutorial or permission to pool different outcomes, time points or duplicate studies.
A concrete task with RigorMed
Turn a forest plot into an evidence-based results interpretation
These are suggested task inputs and outputs to request, not a new execution, a customer outcome or a guarantee of completion.
Materials to provide: A lawfully accessible paper, the plot's results table in readable text, the effect measure and the figure legend.
Requested outputs / delivery goal: An effect-measure checking table, an interpretation draft and a list of unresolved questions.
An example request
Using only the readable text results table and legend I provide, explain each row's effect direction, 95% confidence interval and null value. List any study design or synthesis method that cannot be established separately, and draft an interpretation for a research-group presentation. Do not infer numbers from the image that I have not supplied.
What still needs review
Check the null value of 1 for ratio measures and 0 for difference measures. Whether an interval crosses it does not automatically establish clinical value, causality or no effect.
Read in five steps, not just the diamond
Identify the question and row unit
Check population, comparison, outcome and time point. Determine whether rows are studies, models or subgroups. Check multiple reports of one study, overlapping populations and analysis types in the original text; row count is not automatically the number of independent studies.
Confirm measure, null value and scale
Ratio measures such as RR and OR usually have a null value of 1; mean difference (MD) and standardized mean difference (SMD) usually have 0. Ratio plots often use a logarithmic axis: read the ticks rather than infer differences from equal visual distances. Relabeling different measures does not make them poolable.
Check direction before magnitude
Establish whether the outcome is beneficial or harmful, the comparison order and scale direction. An RR below 1 does not always mean benefit; an MD’s sign depends on units and group order. Verify left/right labels against the original definitions.
Read estimate, interval and weight together
In a conventional meta-analysis plot the square marks the point estimate, the horizontal line usually a 95% confidence interval, and square area the weight under the selected model, not a quality score. Check the actual confidence level in the legend. Do not vote-count crossing rows or treat larger squares as lower-risk-of-bias studies.
Check the synthesis model and scope
The bottom diamond’s centre conventionally marks the pooled estimate and its width usually the confidence interval; follow the legend. Check fixed-effect or random-effects methods, comparability and risk of bias. A random-effects summary describes an average effect, not a guarantee of the same direction in all studies or populations.
Reverse the reference group and the interval together
Write the original comparison as “A relative to B” or “A−B” before reading which side of the line contains the point. The two rows below are independent synthetic teaching examples, not paper results, a product run or a new statistical analysis. CI means 95% confidence interval throughout.
| Original comparison and values | After reversing A/B only | Reason to check |
|---|---|---|
| A relative to B: OR 2.00; CI [1.25, 3.20] | B relative to A: OR 0.50; CI [0.3125, 0.80] | Take the reciprocal of a positive ratio: [L, U] becomes [1/U, 1/L]; reverse the endpoints too. |
| A−B: MD −2.0; CI [−3.5, −0.5] | B−A: MD +2.0; CI [0.5, 3.5] | Negate a difference: [L, U] becomes [−U, −L]; keep the same units. |
Check the conditions first: keep the outcome definition, time point, sample, effect measure and model (including adjustment variables) unchanged; reverse only the comparison order. In the OR example, 2.00×0.50=1. The old upper bound 3.20 maps to the new lower bound 0.3125, and the old lower bound 1.25 to the new upper bound 0.80. Changing only the point estimate, without its interval or left/right labels, creates a contradiction.
Swapping groups is not changing “event” to “non-event”. With risks of 0.20 and 0.10, for example, the RR is 2.00; the non-event RR is 0.80/0.90≈0.889, not 0.50. This is a separate teaching calculation too. Do not interpret an OR as an RR, or interchange OR, RR, HR, MD and SMD by relabelling them.
Units are part of the estimand too: record whether a continuous exposure is per 1 unit, per 10 units or per 1 standard deviation, the MD scale units and what a higher score means. Do not multiply a per-1-unit ratio by 10 to obtain a per-10-unit ratio, or treat rows with different units or adjustment sets as the same result. This section does not refit models or convert exposure increments.
Build a checking table with one estimate per row: source page/table/row; outcome and time; effect measure; comparison/reference groups; exposure increment or scale unit; estimate and CI; model/adjustment set; transformation and unresolved items. Retain original values and the reason for any transformation. Mark missing fields “not reported”; do not extract guessed numbers from the image.
Add to the suggested task request above: “Organize my lawfully usable text results table using these fields. Flag direction or unit mismatches only, retain original numbers, and provide conditional transformation explanations and questions for review. Do not invent results or run a new pooled analysis.” Check each item against the source before using the interpretation in a meeting or manuscript. Consistent directions do not establish poolability, causality or clinical importance.
Four common graphic elements
These meanings apply to clearly labelled conventional meta-analysis plots. Other graphics require their own legend and analytical methods.
| Element | Usual meaning | Does not establish |
|---|---|---|
| Vertical reference line | The measure’s null value, often 1 or 0 | Every measure uses 1, or the left side always favours treatment |
| Square and area | Point estimate and study weight in this synthesis method | A larger area means higher quality or stronger causal credibility |
| Horizontal line | The stated confidence interval, often 95% | Including the null proves groups are the same |
| Bottom diamond | Pooled estimate and its interval | A new-study prediction interval, or effectiveness in every study |
If measure, units, confidence level or model are missing, retain the gap and check the source. Do not ask AI to fill them from appearance.
Synthetic teaching values: examine the range, not just crossing
Suppose an RR is 0.80 with 95% CI 0.60–1.06. The interval contains the null value 1 and is compatible with ratios below and slightly above 1. It does not prove no effect or automatically demonstrate benefit; check outcome and comparison direction. These are teaching values, not a real study or a statistical recalculation.
A separate teaching example is MD −2.0 with 95% CI −3.5–−0.5, excluding 0. Units, group order, an important-difference criterion and risk of bias are still needed to discuss practical meaning. It cannot be directly compared or pooled with that RR, and non-crossing does not establish causality.
Heterogeneity and prediction intervals are not footnotes
Interpret variation alongside populations, interventions or exposures, outcomes, methods and effect directions. I² and τ² are uncertain with few studies: one threshold must not automatically select a model, authorize pooling or grade quality. An unremarkable heterogeneity test does not prove identical effects.
A random-effects summary confidence interval describes uncertainty in the average effect, not the full spread of study effects. Read a reported prediction interval separately with its methods and assumptions. It helps express potential underlying effects across similar studies; few studies or inappropriate distribution assumptions warrant caution. It does not promise an individual patient’s or new study’s result.
Five checks before writing the interpretation
Link the graphic to original text and formal outputs:
- Are row units, study identities, outcome, comparison and time point explicit?
- Do measure, null value, axis, units and direction agree?
- Are estimates, confidence levels, weights and synthesis models sourced?
- Are heterogeneity, bias and applicability explained beyond significance labels?
- Do text, tables and legend agree, with illustrative values clearly marked?
With only an abstract, low-resolution image or incomplete legend, specify unverified information. Partial plot reading is neither full-text appraisal nor clinical advice.
Frequently asked questions
Does crossing the reference line mean no effect?
It does not prove no effect. The interval is compatible with the null and other values; interpret estimate, precision, important differences and limitations without relying only on binary significance.
If one subgroup crosses and another does not, do subgroup effects differ?
That does not establish a difference. Compare subgroups with an appropriate difference or interaction test for the prespecified question; different within-group significance is not a between-group effect difference.
Does a forest plot or pooled diamond mean a high-quality review?
No. Check search, selection, extraction, risk of bias, comparability and synthesis methods. A clear graphic does not show those steps were performed correctly.
Verified sources
The official sources below were checked on 2026-10-09 / RigorMed: 2026-10-13 / 2026-10-22. Access conditions and update dates are rechecked before publication.
- Cochrane Handbook Chapter 10: analysis and meta-analysis
Elements, weights, heterogeneity, subgroup comparisons and prediction intervals, in the intervention-review context.
- Cochrane Handbook Chapter 15: interpreting results
Interpretation of effects, confidence intervals, statistical and practical importance.
- Cochrane Handbook Chapter 6: choosing effect measures and computing estimates
Sections 6.1.2, 6.4.1 and 6.5.1 cover comparison order, ratios/differences, event definitions and scale direction. The numerical transformations here are independently derived synthetic teaching examples, not study results from the chapter.
Start with a specific task
Turn a forest plot into an evidence-based results interpretation
Bring legally usable materials and this page's example request to the workspace. Check the pre-execution material and credit notices before confirming, then review the resulting outputs.