Year: 2014
Type: article
Abstract: Wing et al. (1) recently examined whether behavioral outcomes in the Look AHEAD (Action for Health in Diabetes) intensive lifestyle intervention clustered within treatment groups. Look AHEAD provides a terrific opportunity to examine this issue considering its large sample (2329 participants, 209 intervention groups, and 15 sites). Indicators of adherence showed small to moderate clustering (maximal intraclass correlation coefficient [ICC] of 0.118), whereas weight loss outcomes did not exhibit clustering (ICC of 0.007). Based on these findings, the authors concluded that future weight loss studies need not account for clustering of outcomes in sample size estimation and statistical analysis. We offer several reservations about generalizing this conclusion to future weight loss trials. We recently simulated the impact of clustering effects in behavioral intervention studies with continuous outcomes for an unbalanced design in which groups were only used in the intervention arm, but not in the control arm (2). This is a common design in behavioral weight loss intervention trials, and similar to that of the Look AHEAD trial (3). Our research shows that even moderate clustering can have a significant impact on Type I error. The cluster unadjusted t-test provides poor control of Type I errors with actual empirical significance levels substantially higher than the assumed significance levels (0.05 and 0.01), and as the ICC increases, the empirical significance levels from the unadjusted t-test depart even further from the assumed significance levels. However, the cluster adjusted t-test has actual empirical significance levels that are indistinguishable from the assumed significance levels (Table 1). In addition, if the ICC is ignored when the study is planned, there will be insufficient power to detect outcome differences between intervention arms if in fact the ICC is nonzero. Based on the unique characteristics of the Look AHEAD study, other reasons to consider impact of clustering when designing future weight loss trials are: 1) Look AHEAD was a multicenter trial involving 15 sites and about 11% of variance between treatment groups was attributable to clinic sites. Therefore, group clustering may be more pronounced in single-site trials; 2) Look AHEAD had a number of methodological features that may have reduced its susceptibility to clustering, for example, a highly protocolized and intensive intervention; and 3) the discrepancy between the magnitude of clustering observed for behavioral adherence and weight loss outcomes in Look AHEAD should give a behavioral weight loss researcher pause. One interpretation by the authors for this discrepancy is that any effect of clustering is masked by physiological or behavioral processes that influence weight, but are independent of the effects of the group-based intervention being evaluated. Thus, to ignore clustering in trial planning is to discount a commonly hypothesized mechanism of action for many group-based interventions. Despite the richness of the Look AHEAD data and the robust nature of the findings presented by Wing et al., we recommend that studies continue to consider the potential effect for clustering in power estimation and statistical analysis planning, calculate the magnitude of observed clustering at the conclusion of the trial, and determine whether this should be accounted for in the analysis.
Source: Obesity
DOI: 10.1002/oby.20763
Institution: Rush University Medical Center
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W2136046583