How long does data extraction actually take for a systematic review?

Every review protocol allocates time for screening. Almost none of them budget honestly for extraction, which is where timelines die. Here’s the arithmetic worth doing before your protocol promises a submission date.

The per-paper reality

Suppose a disciplined extractor averages 2–3 hours per included trial. That’s not reading time. It’s finding time: locating the right table among six, the right arm among four, the right timepoint among five, deciding whether “12 weeks” in Table 2 is the same visit as “endpoint” in Figure 3, and transcribing means, SDs, Ns, and event counts without transposing a digit at hour two of the session.

Now the multipliers the protocol committed you to:

30 included trials × 2.5 h            =  75 hours   (first extraction)
× 2 extractors (the gold standard)    = 150 hours
+ reconciliation (~15 min/paper)      = ~158 hours

At a research-assistant rate of $25/hour, that’s roughly $3,900 of labor for the extraction phase alone, before a single forest plot exists. For a solo PhD student doing both passes themselves, it’s a month of full-time work wedged between teaching and the rest of the thesis.

Where the hours actually go

Timing studies of our own extractions, plus every reviewer we’ve asked, converge on the same split: the minority of time goes to transcribing numbers; the majority goes to locating and deciding: which table, which arm, which timepoint, endpoint vs change score, SD vs SE, ITT vs per-protocol. The transcription is mechanical; the locating is exhausting; the deciding is the only part that genuinely needs you.

That split is the argument for changing the process rather than working longer: the mechanical 90% (locate + transcribe + convert) is automatable with an audit trail, which moves your hours onto the deciding: reviewing values against their sources, adjudicating the odd cases, making the judgment calls that will be defended in peer review.

The question to ask any tool (or any RA)

Not “how fast?” but: “when I’m asked where a number came from, what’s the answer?” Speed without provenance just relocates the work to revision time, when the R2 letter asks about the SD in row 14 and nobody remembers which supplementary table it came from. Whatever your process (human, tool, or both), the artifact that survives is the one where every number can answer for itself.