How to use it
- Paste one row per group (each arm of each study) with the sample size, the median and the quartiles, the range, or both. Blank cells are fine.
- Check the columns under the data box and pick the methods (the defaults are the recommended ones).
- Copy the results into Excel or into the forest plot generator, and use “Methods text” for your manuscript. “R code” gives a script that repeats every conversion with the meta package for R. Small-study effects can then be checked in a funnel plot.
The formulas
- Mean from quartiles (Luo 2018): w·(Q1 + Q3)/2 + (1 − w)·median, with w = 0.7 + 0.39/n.
- Mean from range (Luo 2018): w·(min + max)/2 + (1 − w)·median, with w = 4/(4 + n0.75).
- SD from quartiles (Wan 2014): (Q3 − Q1)/η(n), with η(n) = 2Φ−1((0.75n − 0.125)/(n + 0.25)), tabulated for n ≤ 201.
- SD from range (Wan 2014): (max − min)/ξ(n), with ξ(n) = 2Φ−1((n − 0.375)/(n + 0.25)), tabulated for n ≤ 50.
- SD from both (Shi 2020): a weighted combination of the two, with weight 1/(1 + 0.07n0.6) on the range.
Frequently asked questions
When do I need to convert medians to means?
Meta-analyses of continuous outcomes pool means and standard deviations, but many studies report the median with the interquartile range (IQR) or the range, especially for skewed outcomes such as length of stay. The Cochrane Handbook (section 6.5.2.5) allows estimating the mean and SD from these statistics, with a sensitivity analysis, when the data are not too skewed.
Which methods are used?
Means use the method of Luo et al. (2018), which weights the median against the mid-quartile range and mid-range according to the sample size; the older method of Wan et al. (2014) is available too. Standard deviations use Wan et al. (2014) when only the quartiles or only the range are given, and Shi et al. (2020) when both are given. For small samples the SD uses the tabulated values of Wan et al. (2014), exactly as the meta package for R does, and the results match it.
What if the data are skewed?
All these methods assume roughly normal data. If the median is far from the middle of the quartiles or of the range, CleanFigure flags the row. For clearly skewed data, consider methods that model skewness (for example the quantile estimation or Box–Cox methods of McGrath et al., in the estmeansd R package), or pooling medians directly, and report a sensitivity analysis.
How do I use the results in a forest plot?
Convert each arm of each study as its own row, then copy the Mean, SD and n columns into the Mean / SD / n data of the forest plot generator, one row per study with the two arms side by side. Report in your methods which studies were converted and with which method; the “Methods text” button drafts that sentence.
Is my data uploaded?
No. The calculation runs in your browser; nothing is sent to a server.