Research & PolicyFree
Systematic Review Data Extraction Table
One row per study, and every coded value sits beside the sentence it was coded from and the page that sentence is on.
River reads each included study and builds the extraction table one row at a time. Every coded cell arrives with two companions: the sentence the value was read from, quoted as printed, and the location of that sentence in the paper. The population, comparator, outcome and effect estimate columns each get that treatment. Where the study does not report a field, the cell says not reported and the reason goes into a second document rather than into a guess.
Every result that ranks for this query is a blank form. Cochrane publishes one, so does JBI, and each has a single cell per field. That shape is the problem: the coded value replaces the paper's wording, so a second reviewer opening the sheet can only redo the extraction, never check it. Cochrane's own comparison of extraction tools names the missing capability, linking data items to locations in the report to facilitate checking.
Built for the reviewer coordinating a second extractor, the doctoral candidate whose committee will ask where a number came from, and the analyst assembling an evidence table for a decision memo. Reach for it once the full texts are in a folder. The search strategy pack covers the step before that, the duplicate merge plan stops the same study being extracted twice under two citations, and the synthesis writer turns the finished table into prose that argues. Then the results section writer takes over.
The parenthesis after the mean is not labelled
A paper prints post-test numeracy as 42.1 (3.2) and never says what 3.2 is. It could be the standard deviation, the standard error of that mean, or half a confidence interval, and an SD is the SE times the square root of the sample size. A blank form has one cell for it. You type 3.2 into the column headed SD, the sentence you read is gone, and nothing downstream can tell that a decision was made.
Take an illustrative review of peer tutoring and numeracy with 64 pupils per arm. If 3.2 is the standard error, the true deviation is 25.6, and the study's standard error for the mean difference is 4.669 rather than 0.584. Inverse-variance weighting squares that, so coding a standard error as a deviation multiplies the study's weight by exactly the per-arm sample size. Its share of the pooled estimate goes from 5.7 per cent to 79.3 per cent.
So the extraction table carries three cells where a form carries one: the string as printed, the location it sits at, and the coded value with the conversion shown. That is what makes a check possible, and a check is not optional. Extracting outcome data in duplicate is a mandatory Cochrane standard, and errors turned up in 20 of 34 reviews in the evidence it cites. PRISMA item 10a asks you to report how many reviewers collected data and whether they worked independently.
How it works
Send the studies
The included full texts, the outcome your protocol names, and the fields the table has to carry.
Quote before coding
Each field gets the paper's own sentence and its location first, then the coded value beside it.
Conversions get shown
A standard error becoming a deviation, a median becoming a mean, each with the arithmetic printed.
Reconcile in chat
Disagree with a coding, add the study that arrived late, or ask what the empty cells cost.
What you get
- The extraction table as a Sheet, one row per study and one block per outcome
- Every coded value beside the sentence it came from, quoted exactly as the paper printed it
- The page, table or figure each value was read from, for a checking pass
- Conversions shown as arithmetic, so a standard error turned into a deviation is visible
- A second blank column set for the independent extractor, plus a reconciliation column
- A Doc listing every field that could not be extracted and the reason why
Common questions
My studies report the same outcome five different ways. What does the table do with that?
It gets wider and stays honest about it. Each study keeps its own reported form in the verbatim column, and the coded column holds whatever common measure your protocol asked for, with the conversion shown. Studies that cannot be converted get a row saying so. Five reporting styles become five conversions you can audit rather than one column that pretends they matched.
Can it replace the second independent extractor?
No, and it is built on the assumption that you have one. The sheet ships with a duplicate column set the second extractor fills without seeing the first, and a reconciliation column for the disagreements. Cochrane makes dual extraction of outcome data a mandatory standard, so a tool claiming to remove the second person is selling you a protocol violation.
What happens when a study reports no effect estimate at all?
The row stays, the estimate cells say not reported, and the study lands in the unextractable document with the reason and the location you looked. That list is a deliverable, not a failure. It tells you which authors to email, and it is the evidence for the sentence in your methods about data you could not obtain.
Does it read numbers out of figures?
It reads what the axis and labels state and records the value as approximate, with the figure number in the location cell. It will not digitise a plotted point and present the result as reported data. Where a figure is the only source for an outcome, the row says so, because that changes how much the study should weigh.
How many studies can it handle in one pass?
There is no hard ceiling, and batching is still the right way to run it. Send the studies in the order your screening produced them, keep the field definitions identical across batches, and the sheet stays poolable. The slow part is never the typing. It is deciding what a paper's parenthesis meant, which is the decision the verbatim column preserves.
Is this only for clinical reviews?
No. The column set defaults to population, comparator, outcome, measure and effect estimate because that is what most protocols ask for, and you can replace any of it from the intake form. Education, policy and management reviews use the same structure with different labels. The verbatim column matters more outside medicine, where reporting conventions vary further.
Will it change a number to make studies agree?
Never silently. A conversion appears as arithmetic in its own cell, with the input, the formula and the result, so you can reject it. Harmonising is a judgment about whether two studies measured the same construct, and that judgment belongs to you. The sheet's job is to make sure it was made once, visibly, and not by accident.
Systematic Review Data Extraction Table
Fill in the form and your workspace opens with the work already underway.