meta-analysis of proportion_- problem with showing the 95% CI for the I-squared statistic

[Martha Chipinduro]

I performed a subgroup meta-analysis of proportion using the metan command in Stata ver 17.0 and generated a forestplot. The forestplot shows I² reported with its p-value. However, I need to report it with a 95% CI. The log file does not show the 95% CI. Using the suggestion of just adding the option i2ci at the end of the code produeces an error. Without adding i2ci, the forestplot is produced correctly.
We also ran sensitivity analysis using the Generalised Linear Mixed Model (GLMM) with a logit link and the Hartung-Knapp_Sidik_Jonkman (HKSJ) adjustment. Adding i2ci at the end of the working codes resulted in the same error

***main analysis code begin***
metan discordant_pairs total_household _pairs, pr model(ivhet) transform(fturkey, iv) study(study) by(tbincidence_category) sortby(tbincidence_id) forestplot(astext(40) textsize(100) boxscale(50) spacing(1.2) leftjustify range(0 1) dp(2)) extraline(yes) hetinfo(isq h) i2ci
***main analysis code end***

***begin sensitivity analysis code with GLMM with a logit link***
metapreg discordant_pairs total_household _pairs, model(random) studyid(study) lin(logit) by(tbincidence_category) sortby(tnincidence_id) fplot xlabel(0, o.5, 1) texts(1.6) i2ci
***end of sensitivity analysis with GLMM**

***begin sesitivity analysis with HKSJ***
sort tbincidence_id study
meta esize discordant_pairs total_household_pairs, esize(logitprop) zerocells(0.5, only0) studylabel(study)
meta summarise, random(reml) se(khartung) subgroup(tbincidnece_category)
meta forestplot, subgroup(tbincidence_category) transform(invlogit) nullrefine(lcolor(red) lpattern(dash)) xlabel(0 0.5 1, labsize(vlarge)) xsize(9) ysize(12) graphregin(margin(b=25)) i2ci
***end of HKSJ sensitivity analysis***

***begin of error***
option i2ci not allowed
Error in forestplot
r(198);
***end of error***

* Example generated by -dataex-. For more info, type help dataex
clear
input str27 study int(discordant_pairs total_household_pairs) str6 tbincidence_category byte tbincidence_id
"Behr et al., 1998"            9  34 "<20"    1
"Bennett et al., 2002"        80 260 "<20"    1
"Blanco-Guillot et al., 2018"  6  29 "20-100" 2
"Borrell et al., 2009"         4  11 "20-100" 2
"Cavany et al., 2018"         12  44 "<20"    1
"Chen et al., 2022"            5  11 "20-100" 2
"Colangeli et al., 2020"      13  41 "20-100" 2
"Conceicao et al., 2018"       8  36 "20-100" 2
"Dale et al., 2022"            4  16 "<20"    1
"Dheda et al., 2017 "          8  21 ">100"   3
end

[Leonardo Guizzetti]

I don't see that -i2ci- is a documented option for either -metan- (SSC) or the official -meta-. Was this an AI hallucination?

I tend not to find a confidence interval for I² particularly useful when the effect size is a single proportion as they always seem to be too wide to be useful, especially in small meta-analyses as is the case here. If you insist on reporting its confidence interval, you will need to do it by hand. Accessible formulas to do this can be found in Thorlund et al (2012).

Thorlund K, Imberger G, Johnston BC, Walsh M, Awad T, Thabane L, Gluud C, Devereaux PJ, Wetterslev J. Evolution of heterogeneity (I2) estimates and their 95% confidence intervals in large meta-analyses. PLoS One. 2012;7(7):e39471. doi: 10.1371/journal.pone.0039471. Epub 2012 Jul 25.

[Martha Chipinduro]

A true AI hallucination!

Thank you Leonardo Guizzetti. We have 30 studies altogether. I agree that the confidence interval for I² tends to be wider for a single proportion. Thank you for the citation.

[Tiago Pereira]

Hi, Martha.

The use of a 95% confidence interval for I² does not appear to be particularly informative or useful. Overall, I² is not a suitable metric for meta analysis of proportions, as it is typically inflated due to relatively small within study variances. A more appropriate approach is to report 95% prediction intervals, which reflect the expected range of proportions across different populations and better capture between study heterogeneity.

[David Fisher]

Hi Martha et al,
I agree with others that a 95% confidence interval for I² is probably not helpful. However, as the maintainer of the metan package, I should point out that confidence intervals for I-squared should nevertheless be displayed on screen, beneath the text "Heterogeneity measures, calculated from the data". There is indeed no option (such as "i2ci" ) to have this information directly displayed on the forest plot; but metan's forest plots are designed to be flexible, and to incorporate any information you desire, with a little work (see help forestplot).
Best wishes,
David.