Cluster-randomized crossover trial of chlorhexidine-alcohol versus iodine-alcohol for prevention of surgical-site infection (SKINFECT trial).

We read with interest the work of Charehbili and colleagues1, which ‘aimed to investigate whether there is a superiority of chlorhexidine–alcohol over iodine–alcohol for preventing SSI’.

This cluster‐randomized crossover trial was conducted in five hospitals and 3665 patients were included. The authors found that the incidence of surgical‐site infection (SSI) was not different between the groups: 3·8 per cent among patients in the chlorhexidine–alcohol group versus 4·0 per cent in those in the iodine–alcohol group (odds ratio 0·96, 95 per cent c.i. 0·69 to 1·35).

We commend the authors for performing this interesting study, as these results are useful for the choice of the most appropriate preoperative antiseptic. However, we have several statistical suggestions and queries that we would like to communicate to the authors.

The authors concluded that ‘Preoperative skin disinfection with chlorhexidine–alcohol is similar to that for iodine–alcohol with respect to reducing the risk of developing an SSI’. This may be due to an underpowered study.

In fact, sample size was estimated by simulation. Although this approach is efficient, the authors do not provide enough details on the parameters they used. Thus it is not easy to replicate calculations. As mentioned by the CONSORT statement2: ‘the reports of cluster randomized trials should state the assumptions used when calculating the number of clusters and the cluster sample size’.

The authors mention R software for the simulations that led to the final estimation of sample size. But several R packages are available and one may suppose that a package such as clusterPower was used3. Not knowing which package was used does not permit the analysis to be replicated. Moreover, algorithms used may vary between packages and lead to different estimations of sample size.

In a cluster‐randomized crossover trial, a sequence of interventions is assigned to a cluster (group) of individuals. Each cluster receives each intervention in a separate period of time and this leads to ‘cluster periods'4. There is usually a correlation between patients in the same cluster. In addition, within a cluster, patients within the same period may be more similar to one another than to patients in other periods5.

In a cluster‐randomized crossover trial, the sample size estimated by not taking into account the above‐mentioned features must be multiplied by a defined inflation factor. The latter can be approximated by (1 + (n − 1)ρ) − η6, 7, where n is the average number of patients in a cluster during one of the periods, ρ is the intraclass correlation (ICC), and η the interperiod correlation. See, for example, Turner et al.8 or Moerbeek and Teerenstra9 (p. 94) for other approaches to estimate sample size in this context.

Parameters ρ and η can be retrieved from literature or estimated using assumptions or approximations10 (p. 203), for instance: the logarithm of the ICC can be approximated by the logarithm of the prevalence of disease (here, the SSI rate)11; the interclass correlation is intrinsically lower than the ICC12.

Data were analysed using a multilevel model, which is appropriate. Treatment period was considered as a fixed effect and hospitals as random effect. Treatment period could also be considered as a random effect. In their simulations, Morgan et al.5 actually demonstrated that ‘hierarchical models without random effects for period‐within‐cluster, which do not account for any extra within‐period correlation, performed poorly with greatly inflated Type I errors in many scenarios’.

The authors did not report variance components of outcomes: within‐ and between‐participant variance, the ICC, as recommended by some authors13.

In a cluster‐randomized crossover trial, three components of variation are available: variation in cluster mean response; variation in the cluster period mean response; and variation between individual responses within a cluster period4. Small changes in the specification of the within‐cluster–within‐period correlation, or the within‐cluster–between‐period correlation, can increase the required number of clusters4. Thus, as the above‐mentioned correlation parameters were not reported by Charehbili et al.1, the number of clusters required may be larger than that used in the study.

A simulation study showed an association between an increase in cluster size variability and a decrease in statistical power14. The authors did not address this point.

In summary, the results of this study are interesting, but readers should interpret them with caution, according to the statistical methods used for design and analysis of cluster‐randomized crossover trials.

Disclosure

The authors declare no conflict of interest.