How to Do a Meta-Analysis:Pooling Is Not Collecting and Averaging

First, the context. A meta-analysis isn't a standalone task — it usually refers to the statistical pooling done at the final stage of a systematic review, where the question, search, selection, and appraisal were designed in advance.

Which means the quality of a meta-analysis is largely decided before any pooling happens. If the criteria for selecting studies were flawed, no amount of computational precision makes the result trustworthy. (For how the two relate, see our article on systematic reviews versus meta-analyses.)

A meta-analysis question is narrowed to a single PICO (Patient, Intervention, Comparison, Outcome): does this intervention improve this outcome in this population?

Too broad a question mixes studies of fundamentally different kinds, and no later step can justify pooling them. Whether pooling will be defensible is largely decided by this one sentence.

Design a search query, search multiple databases with records kept, and screen against predefined criteria. This is the standard systematic review process, so we'll leave the details to our articles on review design and screening.

AI summaries can assist the first-pass reading at this stage, but inclusion decisions and exclusion reasons are recorded by a human working from the original text.

From the included studies, extract the numbers pooling requires — participant counts, per-group results, effect size data — in a uniform format.

In a meta-analysis, confirming that outcome definitions, measurement time points, and analysis populations correspond across studies is the heart of data extraction. Skip it and you commit the classic error of averaging different things. The practical details are in our article on data extraction for meta-analysis.

With the data assembled, resist the urge to pool immediately. First ask: can these studies actually be combined? Are the populations, interventions, and outcomes similar enough? Are the designs compatible?

"Can be calculated" and "can be combined" are different things. Averaging apples and oranges doesn't produce a meaningful fruit. Whether to pool, and how (fixed versus random effects, for instance), involves statistical judgment — consulting a statistician or supervisor is essential.

Pooled results come with an assessment of between-study variability — heterogeneity. When it's large, asking why the studies differ often yields more important findings than the pooled value itself.

Interpreting heterogeneity statistics and designing subgroup and sensitivity analyses is team work, not solo work. The conclusion of a meta-analysis isn't a single number — it's the number, plus how similar the studies were, plus why they differed.

Finally, return to the question. Is the pooled effect clinically meaningful? How much does the risk of bias in the included studies qualify the conclusion?

This part can't be delegated to AI. Only someone who has read the originals and understands their limitations can write an honest interpretation.

Across this long journey, Paperfy serves as the place you read, verify, and return. Each included paper gets one page with its AI summary, figures, tables, and original PDF. Summaries can be edited by hand, so you can add extraction confirmations or concerns about poolability as notes. It doesn't replace statistical software (R, RevMan) or a citation manager — it supports the reading that comes before them.