Intracluster correlation coefficients and reliability of randomized multicenter stroke trials within VISTA

Abstract

Background: Reliable estimates of intracluster correlation coefficients (ICCs) for specific outcome measures are crucial for sample size calculations of future cluster randomized trials. ICCs indicate the proportion of data variability that is explained by defined levels of clustering.

Aims: In this manuscript, we present potentially valuable and reliable estimates of ICCs for specific baseline and follow-up data.

Method: ICCs were estimated from linear and generalized linear mixed models using maximum likelihood estimation for common measures used in stroke research, including modified Rankin Scale (mRS), National Institutes of Health Stroke Scale (NIHSS), and Barthel Index (BI).

Results: Data were available for 11 841 patients with ischemic stroke from 11 randomized trials. After adjusting for age, thrombolysis, and baseline NIHSS, the median ICC for follow-up data, using center as the level of clustering, ranged from 0·007 to 0·041. The ICCs using trial, continent or year of enrollment as level of clustering were distinctly lower. Less than 1% of the variability of mRS, NIHSS, and BI was explained by any of these three cluster levels.

Conclusion: This compendium of relevant ICC estimates should assist trial planning. For example, the sample size for a cluster trial with 150 patients per center using ordinal analysis of mRS should be inflated by 2·0 due to the ICC of 0·007; whereas the ICC of 0·031 using mRS dichotomized above mRS 0-1, requires inflation by 5·6. The low contribution of trials, year or continent of enrollment to overall variation in outcome offers reassurance that analyses using pooled data from multiple trials in VISTA are unlikely to suffer from bias from these sources.

Keywords: acute; cluster randomization; design effect; intraclass correlation; secondary care; stroke.