How to Extract Data From Graphs for Meta-analysis

When should you extract data from a graph?

Graph digitisation is a fallback method for recovering numerical results that are visible in a figure but not available in a more direct numerical form. The Cochrane Handbook notes that review authors may need to extract data from figures using software when the required numerical data are not available elsewhere, and that software-based extraction is preferable to manual estimation with a ruler.¹

The practical hierarchy is straightforward: first look for the values in the main text, tables, supplementary material, source-data files, trial registries or repositories. If key values remain unavailable, consider contacting the study authors. Digitise the graph when the figure is the only practical source—or when author-provided data cannot be obtained in a reasonable timeframe.

Before digitising: define the data you actually need

It is easy to extract more numbers than a review needs. Before opening digitisation software, define the outcome, comparison, time point, effect measure and summary statistics required by the review protocol. Cochrane recommends planning data collection in advance and collecting sufficiently detailed, unambiguous information to represent the source faithfully.¹

For continuous outcomes, a meta-analysis commonly needs a central estimate such as a mean together with a measure of dispersion or precision, often an SD, SE or confidence interval, plus the sample size. Other review questions may require medians and ranges, proportions, event rates, correlations or individual XY points. Do not assume that every visible element in the graph is relevant to the planned synthesis.

What can be extracted from a scientific graph?

A calibrated graph can provide numerical positions for central values, error-bar endpoints and XY coordinates. Whether those values are suitable for meta-analysis depends on what the graph represents and whether the surrounding paper gives enough information to interpret it correctly.

A reproducible graph-extraction workflow

The mechanics vary between digitisation programs, but the underlying process is similar: identify the figure, define the scale, map pixels to scientific values, click the plotted locations, and preserve enough context to check the result.

1. Identify the source. Record the study ID, article/report, figure number and panel before extracting values.
2. Select the relevant plot. Work from the clearest available source—preferably the original PDF or high-resolution figure rather than a screenshot.
3. Identify the axis type. Determine whether each numerical axis is linear, logarithmic or reversed.
4. Calibrate using labelled reference values. Use well-separated tick marks where possible and ensure the calibration covers the part of the axis containing the values of interest.
5. Extract the central value and uncertainty separately. Follow a consistent marker order so the central estimate is not confused with an endpoint.
6. Add metadata. Record outcome, units, group, time point, sample size, statistic type and error type alongside the number.
7. Check the extraction against the figure. Revisit implausible values, edge cases and short error bars before importing them into the analysis dataset.
8. Preserve provenance. Keep the source reference and, where possible, calibration/audit information with the extracted data.

Studies comparing software-assisted graph extraction with manual visual estimation have found software to be faster and to provide better reliability or accuracy than manual approaches.^2,3 Evaluations in systematic-review settings also support graph digitisation as a practical method when source values are otherwise unavailable.^4,5

Error bars: identify what they mean before extracting them

The visual appearance of an error bar does not tell you whether it represents an SD, SE, confidence interval, IQR or another quantity. That information must come from the figure legend, methods, supplementary material or source data.

This distinction matters because the same graphical distance can represent very different statistical information. If a meta-analysis needs an SD but the figure reports an SE or confidence interval, extract what is actually plotted first and perform any conversion explicitly afterward using the appropriate formula and sample size.

Symmetric and asymmetric uncertainty

If upper and lower error bars are intended to be symmetric around a central estimate, a digitisation workflow may represent the uncertainty as a single magnitude. If the interval is genuinely asymmetric, retain the lower and upper bounds separately rather than forcing them into one symmetric value.

Take special care with logarithmic axes

On a logarithmic axis, equal distances represent equal ratios rather than equal absolute differences. A point halfway between 1 and 100 on a log10 axis is not 50.5; it is 10. The calibration therefore needs to be performed in logarithmic space.

When using Plot Data Extractor Pro, select Log10 and enter the actual values represented by the tick labels—for example, 1 for 10⁰, 10 for 10¹, and 100 for 10². The logarithmic transformation is then handled internally.

How should extracted graph data be checked?

Graph extraction introduces an additional measurement step, so verification should be planned rather than treated as an optional final glance. Cochrane recommends independent duplicate extraction of outcome data in systematic reviews, with an explicit process for resolving discrepancies.¹

At minimum, compare extracted values against the visual source and check that their magnitude, ordering and uncertainty make sense. For important or difficult figures, independent repeat extraction can reveal calibration or marker-placement errors that would otherwise pass unnoticed.

Published evaluations generally report high agreement between graph-extracted and source values, but extraction error can vary by figure and can be larger for uncertainty measures than for central estimates.^4,5 When a borderline or influential result depends on approximate graph-derived data, a sensitivity analysis using plausible alternative values can help assess whether the review conclusion is robust.⁵

Preserve provenance and report graph-derived data transparently

Reproducibility depends on keeping more than the final number. The extraction record should make it possible to determine where the value came from and how it was obtained. Guidance on reproducible meta-analysis data extraction similarly emphasizes retaining transparent, shareable extraction information.⁶

In the review methods, state that numerical data were digitised from figures where direct numerical values were unavailable, identify the software used, and describe any duplicate extraction, verification or sensitivity-analysis procedure that was applied.

Using Plot Data Extractor for meta-analysis data extraction

Plot Data Extractor provides two workflows. The free Chrome extension is designed for point-and-click extraction from plots displayed in Chrome. Plot Data Extractor Pro provides a dedicated PDF workflow for evidence synthesis, including study/figure metadata, linear and log10 calibration, central and uncertainty extraction, an evidence table, local project files, and separate analysis and audit CSV exports.

Plot Data Extractor Pro processes research PDFs and extracted scientific values locally in the browser. Its extraction methods have been validated using controlled figures and independently published scientific plots. See the dedicated Methods & Validation resource for the calibration equations, validation design, results and limitations.

Related: Plot Data Extractor Pro User Guide · Plot Data Extractor: Methods & Validation