How do I meta-analyse prevalence?
Pool proportions on a transformed scale rather than raw, because a proportion's variance depends on the proportion itself and the usual approximations break down near 0% and 100%. A logit transformation, or better a generalised linear mixed model fitted to the raw numerators and denominators, handles this properly. Use random-effects — prevalence is almost always genuinely heterogeneous.
Why raw proportions misbehave
A proportion is bounded between 0 and 1, and its variance is a function of its own value — largest near 0.5, vanishing as it approaches either bound. That breaks the assumptions behind ordinary inverse-variance pooling.
The practical symptoms are familiar: confidence intervals that extend below 0% or above 100%, and studies reporting zero events receiving effectively infinite weight or being dropped altogether. Neither is a rounding problem; both are the model telling you the scale is wrong.
Which transformation
- Logit
- Well understood, back-transforms cleanly, and behaves sensibly across most of the range. A reasonable default. Studies with zero or total events still need care, since the logit is undefined at the bounds.
- Generalised linear mixed model
- Models the numerators and denominators directly with a binomial likelihood and a random effect, rather than transforming a summary statistic. It handles sparse and zero-event studies without continuity corrections, and it is the better choice when you can fit it.
- Double-arcsine
- Widely used and stabilises the variance, but back-transforming the pooled estimate is known to be problematic when study sample sizes vary — the reverse transformation is not well defined by a single number, and it can produce estimates outside the plausible range. Prefer one of the above.
Expect heterogeneity, and do not apologise for it
Prevalence varies by geography, period, case definition, and how hard anyone looked. A prevalence meta-analysis will usually report a very high I², and — per the guide on that statistic — this is close to uninformative. It is telling you that real populations differ, which you knew.
The useful work is explaining the variation rather than summarising it away. Pre-specify subgroups (region, diagnostic criteria, sampling method, year) and report those, and treat the single pooled number as secondary to the pattern.
If your studies used different case definitions, a pooled prevalence averages estimates of different things. Check definitions before pooling, not after.
Where this answer stops
A pooled prevalence is an average across whatever populations happened to be studied, weighted by sample size rather than by any sampling frame. It is not an estimate of prevalence in any real population unless the studies were designed to represent one, and they rarely were.