What a data extraction form for a systematic review should contain

Most extraction forms collect enough to fill a results table and not enough to defend a single number. The difference shows up months later, when a co-author asks where a figure came from and the honest answer is “I think it was Table 2.” A good form is designed backwards from the errors it has to catch. Four fields do most of that work.

Arm-level N, at the timepoint you are extracting

Not the randomized N. The number analyzed in that arm, at that timepoint, for that outcome. These drift apart the moment there is dropout, and effect-size arithmetic uses the analyzed count. Suppose 200 were randomized 1:1, but the 12-week outcome has 84 per arm after loss to follow-up, with 24 events in one arm:

p using randomized N = 24 / 100 = 0.240
p using analyzed N   = 24 /  84 = 0.286

That gap propagates straight into every risk ratio and risk difference downstream. One N field, captured per arm and per timepoint rather than once per study, closes it. Continuous outcomes have the same failure: an SD is only meaningful next to the N it was computed from.

Measure, unit, and direction as three separate fields

“Change in depression score” is not an extractable value. You need the instrument (which HAM-D, which subscale), the unit (raw points, percent change, log scale), and the direction of benefit (does lower mean better). Direction is the one people fold into their heads instead of the form, and it is where signs flip. If one trial reports a scale where a negative change is improvement and another where positive is improvement, pooling them without a direction field silently subtracts trials that agree from trials that disagree. Record direction explicitly and the sign convention becomes a data question, not a memory test.

An SD-versus-SE flag on every dispersion value

A number in parentheses after a mean is worthless until you know what it is. SD, SE, a 95% CI, and an IQR are four different quantities that look nearly identical on the page, and they differ by factors that scale with N. Mistaking an SE for an SD at n=25 per arm understates the SD by about five-fold. So the dispersion field should never be a bare number. It should be a value plus a type tag. The tag is also what tells you a conversion is owed later, and which one (Handbook §6.5.2 for SE and CI; Wan 2014 for IQR).

The field everyone omits: source location, per value

Most forms have one “source” or “notes” column per study, if that. What they need is a location attached to each value: “Table 3, row 2, 12-week column” for this mean, “text p.7 para 2” for that N. It feels redundant while you are looking right at the PDF. It is the single field people most regret skipping, for three reasons.

First, verification. A second checker with a per-value location verifies in seconds; without one they re-hunt the whole paper. Second, correction. When a value looks wrong at analysis, a location tells you whether to re-read one cell or re-extract the study. Third, defense. “Table 3, adjusted model, ITT population” is a complete answer at peer review two years later; “it was in there somewhere” is not. Per-value provenance is what turns an extraction from a claim into a record.

The habit

You can add all four to a Covidence form or an Excel template today: arm-level N per timepoint, measure/unit/direction split apart, a type tag on every dispersion value, and a location string beside each number. None of it is exotic; all of it is skipped under deadline. The reason to build extraction on tooling that carries a source location for every value, automatically, is that it is the field you will never regret having and always regret omitting. It’s also the one no one reliably fills in by hand at 11pm.