Retrieve a Missing Standard Deviation Without Guessing

Search before reconstructing

A missing SD in the main results table may be available in a supplement, companion report, registry result, correction, or author response. Search those sources using the exact outcome, group, time point, and analyzed sample size. Record where you looked and which version of the report you used.

Keep a structured request ready for the authors: identify the outcome and visit, then ask for the group mean, SD, and analyzed number. A general request for “more data” makes it harder for someone to supply the right summary. Record unanswered requests without replacing the missing value with zero.

Identify the uncertainty measure

An SD describes variation in observations; an SE describes uncertainty in an estimate. For a simple independent sample mean, SE equals SD divided by the square root of n, so SD can be recovered as SE times the square root of n. NIST's confidence-limit reference shows this relationship in the interval for a mean.

In a fictional group with n = 64 and a clearly reported SE of 1.5 for the unadjusted mean, the recovered SD is 1.5 × 8 = 12. This is a derivation from specified inputs, not an imputed observation. It does not apply automatically to an adjusted model coefficient, a clustered estimate, or a between-group difference.

Preserve the derivation conditions

If you start from a confidence interval, establish its level, construction, and degrees of freedom first. A small-sample mean interval may require a t critical value. Follow the confidence-interval conversion guide before converting, and retain the original endpoints and method.

The continuous-outcome extraction worksheet provides a place to distinguish an SD copied from the report from one calculated using other statistics. Mark the analyzed n explicitly: baseline enrollment is not necessarily the relevant n at follow-up.

Treat estimation as a separate decision

A median, range, or interquartile range does not uniquely determine the SD. Published estimators, including Wan and colleagues' methods, use distributional assumptions and available summary information. Assess whether those assumptions are plausible for the outcome, especially with skewness or bounded scores.

If estimation is necessary under the review plan, label it, cite the method, retain its inputs, and examine sensitivity to that choice. Do not mix reported and estimated values without a provenance field. Finish with an unresolved-data list so missing information remains visible when interpreting the synthesis rather than disappearing inside a spreadsheet.

Sources checked on 30 October 2026; numerical example independently recalculated.