For this analysis and the one below, the prevalence of myopericarditis, which was actually zero in the dataset, was treated as having the lowest possible prevalence on the prevalence scale (0.01 %).
We call these the S+T formulations of the model for true prevalence, scale parameter and shape parameter.
These quantities were then back-transformed to the original prevalence scale.
The observed estimates are very close to the fitted regression lines in the prevalence scale.
Since the regression was performed in the logit scale, its reverse transformation was applied to obtain the estimated values in the prevalence scale.
The prevalence scale assesses the presence or absence (Yes/No) of each symptom with a higher overall score indicating a higher symptom occurrence.
Empirically, the linear assumption between the NCHS and WHO estimates in the logit scale was supported by the observed relationship displayed in their scatter plots along with the fitted lines in the prevalence scale.
Since the prevalence scale is between 0 and 1, to perform the regression in the entire real line, the logit transformation was applied to both estimates before fitting them to the regression model.
Estimates of the median and 5th and 95th percentiles of the distribution of true prevalences (i.e., the distribution of study prevalences that excludes variation due to sampling error) were obtained by back-transforming estimates on the logit scale to the prevalence scale.
The scatter plots of the two prevalence estimates, WHO vs. NCHS, used for the algorithm validation are presented along with the predicted lines and their 95% CI bands in Figure 2. Since the linear regression models were fitted in the logit scale, the 95% CI bands in the prevalence scale have funnel shapes.
Interpreted on the prevalence scale, multivariable meta-regression suggested that prevalence is on average 1.1 % (95 % CI 0.1 to 2.1 %) higher in women than men and 1.1 % (95 % CI 0.0 to 2.2 %) higher in 20 to 24 year olds compared to under 20 year olds.
