Single-group meta-analysis (metamean/metaprop) (Shinyapp)

This app pools single-group (single-arm) data — either means or proportions — when there is no comparator arm, using meta::metamean() and meta::metaprop(). It is deliberately minimal: paste a small table, read the pooled estimate and the forest plot.

How to use

The app opens with example data already in the table, on both the mean and the proportion side, so you can press Run straight away and see what the output looks like before committing your own numbers. To use your own data, paste a tab-separated block over it, for example straight out of Excel; there is no file upload. Once the table is loaded it can also be edited in place, which is convenient for fixing a single mistyped cell without re-pasting everything.

  • Means: columns studlab, n, mean, sd.
  • Proportions: columns studlab, event, n.

The output is a forest plot drawn in the page and a text summary. Fixed-effect and random-effects results are always reported together, so the influence of heterogeneity on the pooled estimate is visible without changing any setting.

Two worked examples, so that you can check the app is behaving as you expect. The three studies loaded by default on the mean side, with (n, mean, SD) of (30, 50, 10), (25, 55, 12) and (40, 48, 9), give a fixed-effect estimate of 49.8764 [47.8837; 51.8691] and a random-effects estimate of 50.6051 [41.9652; 59.2450], with tau-squared 7.2749 and I-squared 68.3%. Three studies of proportions with 12/30, 15/40 and 3/25 events give 0.3158 [0.2304; 0.4157] fixed and 0.2980 [0.0825; 0.6671] random, with I-squared 62.7%.

Method and primary sources

The two analysis paths are fixed. There is no interface for choosing a summary measure or a between-study variance estimator, and knowing exactly what is being run matters more here than in a general-purpose tool, because for single-group data the choice of transformation is itself a methodological decision.

  • Means are pooled on the untransformed (raw) mean scale, corresponding to sm = "MRAW" in meta::metamean(), with the DerSimonian-Laird estimator for tau-squared, the Jackson method for the confidence interval of tau-squared, and the Hartung-Knapp adjustment.
  • Proportions are pooled after a logit transformation, corresponding to sm = "PLOGIT" in meta::metaprop(), fitted as a random-intercept logistic regression (a generalised linear mixed model), with maximum likelihood for tau-squared and a t-distribution confidence interval. Results are back-transformed to the proportion scale for display.

Consequently the Freeman-Tukey double arcsine transformation (PFT), untransformed proportions (PRAW) and the log-transformed mean (MLN) cannot be selected here. If your protocol specifies one of those, run the analysis in R directly. The general case for careful handling of proportions — the behaviour of confidence intervals near 0 and 1, and the consequences of the transformation chosen — is set out in [1]; the random-intercept logistic model used here is the option that avoids the back-transformation problems of the double arcsine approach. Heterogeneity is reported as I-squared [2].

How to cite

Cite the underlying method and state the settings explicitly, because they are not selectable and therefore not obvious to a reader. For proportions:

Proportions were pooled using a random-effects model on the logit scale, fitted as a random-intercept logistic regression with maximum likelihood estimation of the between-study variance (meta::metaprop()), using the Single-Group Meta-Analysis app (Furukawa Y, https://yukifurukawa.jp/single-group-meta-analysis-shinyapp/), accessed [date].

For means, replace the second clause with: pooled on the untransformed mean scale with the DerSimonian-Laird estimator and the Hartung-Knapp adjustment (meta::metamean()).

Limitations

  • No choice of transformation or estimator. The settings listed above are the only ones available.
  • No subgroup analysis, meta-regression, funnel plot, publication-bias test or GRADE assessment. For those, use a pairwise tool or run the analysis in R.
  • No download. The forest plot is drawn in the page and the summary is text on screen; there is no export button, so copy what you need or take a screen capture.
  • No file upload. Your own data have to be pasted in over the example data that the app starts with.
  • The methodological limitation is the design, not the software: pooling single-arm results discards randomisation. A pooled single-group mean or proportion describes a set of study populations under a set of study conditions, and it does not estimate a treatment effect. Between-study heterogeneity in such analyses is usually substantial, and the prediction of what would happen in a new setting is correspondingly weak.

References

  1. Barendregt JJ, Doi SA, Lee YY, Norman RE, Vos T. Meta-analysis of prevalence. J Epidemiol Community Health. 2013;67(11):974-978. PubMed 23963506
  2. Higgins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557-560. PubMed 12958120