Pilot a Data-Extraction Form Before Scaling the Review

Begin with the intended analysis row

List the fields you need to produce one defensible analysis record: study identity, population, comparison, outcome definition, time point, effect measure, estimate, uncertainty, and analysis population. Add source location and extraction status. If an analyst cannot tell which outcome and population a number describes, the row is not complete.

Cochrane's data-collection chapter emphasizes structured collection, linking reports of the same study, and piloting forms. Use the pilot to find missing decisions before a large team repeats them hundreds of times.

Choose reports that stress the form

Include examples with multiple time points, supplementary tables, adjusted and unadjusted models, and at least one missing or ambiguous field. If your review includes unusual designs, include one in the pilot. Selecting only a neatly reported study tests typing more than extraction judgment.

Have two reviewers independently extract the pilot material under the same instructions. Keep their original versions so reconciliation does not erase the evidence of where the form failed. Compare definitions and source passages, not just whether the copied numbers match.

Make missingness and transformations explicit

Use distinct labels for not reported, not applicable, unclear, and not yet checked. A blank cell or numeric zero should not carry all four meanings. Store reported numbers separately from derived quantities, with the conversion method and input values next to the derived result.

For continuous outcomes, the continuous-outcome extraction worksheet can support a pilot record. In a fictional trial, a table might label a value “mean ± error” while a footnote identifies the error as SE. A dedicated uncertainty-type field catches the ambiguity before someone treats the number as an SD.

Reconcile with an audit trail

For each disagreement, record the resolution, reason, source, and whether the instructions changed. Add a field only if it supports a planned decision or check; excessive unused fields increase workload without necessarily improving fidelity. Conversely, do not delete a troublesome field merely because the first pilot report lacks the information.

Confirm that multiple publications share a study identifier while retaining separate report identifiers. Check how corrections and author correspondence will update an existing record without losing provenance.

Finally, export the pilot to the format your analysis actually uses. Check decimal conventions, category labels, units, duplicate rows, and preserved identifiers. A form that looks complete on screen can still produce an unusable analysis file. Freeze the revised version and repeat the affected pilot checks before full extraction begins.

Source checked on 30 October 2026; example is fictional.