Incidence and prevalence are reported in terms of their proportions (discussed above). As described by Barendregt, Doi [
4] this implies two important statistical considerations. The first, is that prevalence/incidence will always fall between the values of zero and one. The second, is that the sum of the prevalence/incidence over different categories should always equal one [
4]. These rules are important when considering the pooling of proportional data to include in a proportional meta-analysis. Without transformation of the included proportional data the accompanying meta-analyses experience two threats to statistical conclusion validity [
35‐
37]. Firstly, the confidence limits fall outside of the established zero to one range [
4]; this may impact on the readability and presentation of the pooled data as a forest-plot. The second concern, and by far the most prudent, is that the variance from studies contributing proportional data at the extreme ends of the zero to one range tends toward zero [
4]. This in turn, artificially inflates the weight that these studies contribute towards the pooled-prevalence estimate. Transformation of that data is therefore required during the meta-analysis process to deal with these problems [
4,
16]. The two most common methods for performing this transformation are the double arcsine transformation (Freeman-Tukey transformation) and the logit transformation [
4]. Both of these transformations calculate a pooled prevalence estimate with a 95% confidence interval under both the fixed and random-effects model. While the logit transformation solves the problem of confidence interval estimates falling outside the zero to one range, it does not necessarily resolve the issues regarding variance from extreme proportional datasets. As the double arcsine transformation addresses both problems listed above it is the preferred transformation method when performing proportional meta-analysis. Once the meta-analysis has been performed on the transformed proportions, a back-transformation is required. For log, logit, and arcsine methods, the back-transformation is straightforward. However, for the Freeman-Tukey double arcsine method there is still no consensus about which back-transformation method should be used [
38]. A detailed breakdown and review of transformation and back-transformation methods can be found in the seminal prevalence meta-analysis work from Barendregt, Doi [
4]. It is important to note that transformation methods may be dispensable when proportions for studies are close to 50%, for example, situations in which an effect or event is expected to be normally distributed.
As with traditional comparative methods of meta-analysis, there are different options in terms of model choice when performing a proportional meta-analysis (discussed above) [
39]. When considering that epidemiological factors typically measured using proportional data are well-known to vary between population characteristics, it has been previously recommended that proportional meta-analyses are performed using the random-effects model [
40]. When considering that the fixed-effect model assumes that there is one true estimate measured across studies, this is unlikely to hold true for proportional data synthesised from multiple independent studies. While performing a proportional meta-analysis using the fixed-effect model is possible, authors should be aware of the assumptions this model has on the data and the subsequent inferences that can be made from the final pooled-prevalence estimate [
16].