To the Editor: In an article in the 1 April issue of the Journal,1 Sturm et al reported on a GP-based stroke audit ("ASAP") and stated that "the information obtained is likely to be representative of most Australian general practice environments". Without further information, we cannot be as confident.

First, their sampling strategy was unconventional. Of all registered GPs from five Australian States and one Territory who were initially approached in May 2000, only 10.2% (n = 1850) of eligible GPs expressed interest in participating in the study. From each of 22 "geographical regions", up to 18 GPs were recruited, initially by random sampling and then by replacement, to obtain a sample of 396 GPs, of whom 321 (81%) provided data. No GP data by State and Territory or "geographical region" were provided to allow readers to judge the possibility of sampling bias. Unpublished data from our own GP survey about stroke issues in New South Wales raise this possibility. We conducted a postal survey of 490 randomly selected GPs from November 2000 to February 2001 (response rate, 60%). None of the 296 participating GPs stated they were enrolled in a stroke clinical audit.

Second, although patients were clustered within GPs, no intracluster correlations (ICCs) were reported. Outcomes (eg, disease morbidity and risk factors) for patients recruited from general practices tend to be correlated at the GP level.2 ICCs quantify the extent to which individuals within clusters (such as a GP's practice) are similar to each other relative to individuals from other clusters. Conventional formulas for calculating confidence intervals assume that the ICC is zero (ie, no clustering). Yet, where correlation within clusters does exist (ie, ICC > 0), the effective sample size is reduced and the associated CIs are inevitably wider. For any given ICC greater than zero, larger cluster sizes also further reduce the effective sample size. Applying appropriate formulas,3 we calculated effective sample sizes for risk factors in the ASAP study, assuming three different magnitudes of ICCs, ranging from relatively modest (0.015) through more substantive (0.1) (see Box). Given the large denominator of the ASAP study, our methodological concern may be only minor in terms of the width of the CIs reported, but the reader is unable to judge whether or not this is the case, as no ICCs were reported. As sample-size calculations for future interventional studies would be informed by publication of ICCs,4 we encourage such reporting in future.

Third, we believe the authors' quantitative findings would have been most useful if they had been age-adjusted in line with Australian community norms.