Sir, the paper by Humphris et al.1 has some excellent examples of how to report statistical data. They include a confidence interval of the difference in state anxiety scores between those given the Modified Dental Anxiety Scale and a control group. This confidence interval crosses zero but does not include the pre-specified value for an important effect of one unit, and so they can legitimately claim that being given the Scale does not increase anxiety.

However, if one looks closely at the methods, one discovers that patients were randomised, not individually, but by whole sessions (to avoid 'contamination'). This design is known as a cluster randomised trial. While this is a good research design to avoid contamination, sadly it means that methods that ignore this aspect of the design, such as the t-test described by the authors, are invalid.2 In practice it is most likely that the reported p-values will be too small, so the conclusion of non-significance still holds. However, the true confidence intervals are likely to be wider than those reported, and so may possibly include the effect size of one unit.

Cluster trials are increasingly being used in dental research. A recent paper found seven out of 43 (16%) trials in dentistry were cluster designs.3 It is important to be aware that they require rather more sophisticated methods of analysis, which are not readily available in some of the more commonly used statistical packages yet.

The authors respond: We are grateful to Professor Campbell for alerting us to the necessity of adopting statistical approaches that control for clustering. She is correct in identifying that our randomisation procedure was by session rather than by individual patient. Expediency dictated that we randomise by session to limit the disruption in the waiting room where data collection took place. In order to investigate for the possible influence of clustering on our reported results we adjusted our outcome variable by the 'design effect' (ICCs) as recommended by Campbell (2000) in her letter. We used ICCs that ranged from 0.05 to 0.01. Our results demonstrated that the z values range from 1.048 to 0.689 - that is, lower than the effect we reported of 1.275 (as correctly predicted by Professor Campbell). The confidence interval, however, may include a one unit difference in the MDAS although readers should be aware that this may be border-line. It is also worth stating that we adopted a small effect size 0.22 in our study thereby providing a limited threshold for the intervention to reach. In summary, we support the call for dental researchers to attend to issues that minimise bias in their findings; however we feel that the balance of evidence from our reported study swings towards the minimal effect of using formal dental anxiety questioning on state anxiety levels in patients attending their dentist.