A practical guide to choosing a meta-analysis topic that is feasible, novel, and publishable, using FINER, PICO, and a feasibility check before you commit.
Most failed meta-analyses do not fail at the statistics stage. They fail at the very first decision: the topic. By the time you discover that only three usable studies exist, or that someone published the same review last year, you have already spent weeks screening abstracts. The topic is the part of the project with the highest leverage and the least attention paid to it.
This guide walks through how to pick a meta-analysis topic that is answerable, that adds something to the field, and that you can actually finish. It is written for graduate students, clinicians, and researchers planning their first or second quantitative synthesis, but the checks apply at any level.
A meta-analysis combines numerical results from multiple independent studies to produce a pooled estimate. That single sentence carries two constraints that rule out a lot of otherwise interesting questions.
First, the studies you want to combine have to measure something comparable. If three trials report a continuous outcome on different scales and four report a binary outcome, you cannot simply pool them without careful conversion, and sometimes you cannot pool them at all.
Second, the results have to be reported in a way that lets you extract or reconstruct an effect size. A paper that reports "the intervention group improved significantly" without means, standard deviations, or event counts is, for your purposes, almost useless.
So a good meta-analysis topic is not just an interesting question. It is an interesting question for which a body of quantitative, comparable, extractable evidence already exists. Hold that idea in mind, because most of the screening below is really about testing it.
A common trap, especially for people who have just learned the technique, is to decide "I want to do a meta-analysis" and then go hunting for any topic that fits. That order tends to produce dull, derivative reviews.
Work the other way around. Begin with a question that someone in your field would actually want answered. Does intermittent fasting lower HbA1c in adults with type 2 diabetes? Does a particular soil amendment reduce pesticide leaching? Does cognitive behavioral therapy delivered online work as well as in person? The method is a tool for answering the question, not the reason for the project.
The test is simple. If you imagine presenting your pooled result at a conference, would a clinician change a recommendation, a policymaker reconsider a guideline, or a researcher redesign their next study? If the honest answer is no, the topic is probably too marginal to be worth a year of your life.
The FINER criteria, borrowed from clinical research design, are a quick filter before you invest any real time.
If a topic fails Feasible or Novel, stop and reshape it. Those two are the ones that waste months.
Once you have a candidate topic, write it out using the PICO structure (or PICOS, adding study design). This forces vague ideas into something you can actually search and screen against.
A topic that resists being written in PICO terms is usually too broad. "The effects of exercise on health" is a research program, not a meta-analysis. "The effect of supervised resistance training versus usual care on six-minute walk distance in adults with heart failure" is a topic. The second version already tells you what to search for and what to extract.
Before you commit, find out who got there first. Three places are worth searching in this order:
Search PROSPERO for registered protocols. A review can be registered and in progress without being published yet, and that is exactly the situation you want to discover early rather than after submission.
Search the Cochrane Library and PubMed (or your field's equivalent database) for "systematic review" and "meta-analysis" alongside your key terms. Limit to the last three to five years.
Skim any recent reviews you find. The goal is not only to see whether your question is taken, but whether the existing reviews are any good.
Finding a published meta-analysis on your topic is not automatically fatal. It is fatal only if the existing review is recent, well-conducted, and comprehensive. If the last synthesis is from 2018 and a dozen trials have appeared since, an update is justified. If the existing review pooled incomparable outcomes, ignored risk of bias, or missed a major database, there is room to do it properly. Just be honest about whether your version genuinely improves on what exists, because reviewers will ask.
You can technically run a meta-analysis on two studies. Whether you should is another matter.
With very few studies, your pooled estimate is unstable, your test for heterogeneity has almost no power, and any random-effects model is estimating the between-study variance from a handful of points, which it does badly. There is no universal minimum, and editors disagree, but as a working heuristic: fewer than about five eligible studies is a warning sign, and any subgroup or meta-regression you are hoping to run needs considerably more than that. If you are eyeing publication bias tests like funnel plot asymmetry, you generally want ten or more.
Do a quick scoping search and count realistically. Not "studies on this topic," but studies that match your PICO, report a usable outcome, and are in a language you can handle. That number is almost always smaller than your first guess.
Heterogeneity is the variation in true effects across studies, and a little of it is normal and even expected. The problem is the apples-and-oranges kind, where you have pooled studies so different that the average is meaningless.
Imagine combining trials of a drug at wildly different doses, in different disease severities, with different follow-up windows. You will get a number, and that number will describe nothing real. A good topic is narrow enough that pooling makes conceptual sense, while broad enough to have the studies to support it. Finding that balance is most of the art in topic selection.
A useful exercise: before you start, predict the main sources of clinical and methodological diversity in your candidate studies. If the list is long and you cannot imagine a clean subgroup structure, the topic may be too sprawling.
This is the practical check people skip, and it sinks more projects than any statistical issue. Pull five or six of the studies you expect to include and look at how they report your primary outcome.
Are the means and standard deviations there, or only p-values and graphs? Are event counts reported, or just relative risks without denominators? Do studies use the same scale, or will you be converting between instruments? Is there enough information to compute or impute what you need?
If half your candidate studies report the outcome in a form you cannot convert, your effective sample shrinks fast. Better to learn that now than after full-text screening.
Run a candidate topic through these questions. If you cannot answer yes to most of them, reshape the topic before going further.
That last point is easy to wave away and worth taking seriously. A brilliant topic you cannot complete is worth less than a modest one you can.
When a topic survives the checklist, write a protocol and register it, usually on PROSPERO. This is not bureaucratic box-ticking. A registered protocol locks in your question, your inclusion criteria, and your planned analysis before you see the results, which is the main defense against the kind of post-hoc decisions that make a review untrustworthy. It also serves as a public claim on the question, so a second team is less likely to duplicate your effort.
Choosing the topic is the decision with the highest leverage, but the work that follows, screening, extraction, effect-size calculation, heterogeneity assessment, and forest plots, is where most of the hours go. This is the part Verflux is built to compress: it manages screening, pulls your data into a structured extraction sheet, and runs the pooled models (Hedges' g, DerSimonian-Laird random effects, Cochran's Q, I², prediction intervals) so you spend your time on judgment calls rather than spreadsheet wrangling.
But no tool can rescue a topic that was never answerable. Get the question right first, confirm the evidence exists, and the rest becomes a process you can run with confidence.
There is no fixed minimum, and you can pool as few as two. In practice, fewer than about five eligible studies makes your pooled estimate unstable and your heterogeneity and publication-bias tests nearly useless. Subgroup analyses and meta-regression need more again. Count realistically during a scoping search before committing.
Yes, if you can justify it. An update is reasonable when the existing review is several years old and many new studies have appeared, or when the previous synthesis had clear methodological problems such as pooling incomparable outcomes or missing major databases. Reviewers will expect you to state plainly how yours improves on what exists.
A systematic review is a structured, reproducible search and appraisal of all studies that meet predefined criteria. A meta-analysis is the optional statistical step that combines their quantitative results into a pooled estimate. Every meta-analysis sits inside a systematic review, but not every systematic review includes a meta-analysis, often because the studies are too heterogeneous to pool.
Write it in PICO terms and tighten each element. Specify the population more precisely, pin down a single comparator, and commit to one primary outcome measured one way. If the question still covers wildly different doses, designs, or settings, split it or pick the slice where comparable evidence actually exists.
Start with PROSPERO for registered and in-progress protocols, then search the Cochrane Library and a database like PubMed for recent systematic reviews and meta-analyses on your key terms. Reviewing what you find tells you both whether the question is taken and whether the existing work leaves room to do better.
All analytical features included in the free trial. No credit card, no installation, no R or Python.
Create free account