Statistical issues in the analysis of non-pharmacological therapy trials with clustering by...
1
INVITED SPEAKER PRESENTATION Open Access Statistical issues in the analysis of non- pharmacological therapy trials with clustering by care-provider or therapy group Chris Roberts From Clinical Trials Methodology Conference 2011 Bristol, UK. 4-5 October 2011 In trials of physical and talking therapies or group admi- nistered treatments, clustering of patients within care- provider or treatment group has implications for sample size and statistical analysis analogous to those found in cluster randomised trials [1]. Between-cluster variation reduces precision and power when estimating treatment effects. Statistical analyses that fail to take account of this form clustering rest on the assumption of no clus- tering effect and may therefore lack conclusion validity. In trials with treatment related clustering, the cluster size may differ systematically between treatment arms, due to differing number of therapists in each arm or dif- ferences in the size of therapy groups between arms. The intra-cluster correlation due to therapist or therapy group may also differ between treatment arms. An extreme case of such heterogeneity is the partially nested design in which the clustering effect is absent in one treatment arm. Where both cluster size and the underly- ing variance components differ between treatments, fail- ure to correctly model heterogeneity can bias estimates of the intra-cluster correlation coefficient and test size [1]. This contrasts with cluster-randomised trials where the distribution of cluster sizes in each treatment arm will be similar due to randomisation making cluster ran- domised trials robust to heteroscedasticity. In a cluster randomised trial, it is generally assumed that each subject belongs to just a single unit of rando- misation so that the design is hierarchical. In non-phar- macological therapy trials patients may receive treatment from more than one therapist or in more than one therapy group so the multilevel model is no longer strictly hierarchical. Statistical analysis will therefore require simplifying assumptions regarding the pattern of clustering such as use of a primary therapist or primary group for each patient or the application of a multiple membership model [2]. In conclusion, analysis of trials with treatment related clustering may therefore require more complex methods of analysis than cluster randomised trials. Published: 13 December 2011 References 1. Roberts C, Roberts SA: Design and analysis of clinical trials with clustering effects due to treatment. Clinical Trials 2005, 2:152-162. 2. Browne W, Goldstein H, Rasbash J: Multiple membership multiple classification (MMMC) models. Statistical Modelling 2001, 1:103-124. doi:10.1186/1745-6215-12-S1-A21 Cite this article as: Roberts: Statistical issues in the analysis of non- pharmacological therapy trials with clustering by care-provider or therapy group. Trials 2011 12(Suppl 1):A21. Submit your next manuscript to BioMed Central and take full advantage of: • Convenient online submission • Thorough peer review • No space constraints or color figure charges • Immediate publication on acceptance • Inclusion in PubMed, CAS, Scopus and Google Scholar • Research which is freely available for redistribution Submit your manuscript at www.biomedcentral.com/submit University of Manchester, Manchester, M13 9PL, UK Roberts Trials 2011, 12(Suppl 1):A21 http://www.trialsjournal.com/content/12/S1/A21 TRIALS © 2011 Roberts; licensee BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
-
Upload
chris-roberts -
Category
Documents
-
view
212 -
download
0
Transcript of Statistical issues in the analysis of non-pharmacological therapy trials with clustering by...
INVITED SPEAKER PRESENTATION Open Access
Statistical issues in the analysis of non-pharmacological therapy trials with clustering bycare-provider or therapy groupChris Roberts
From Clinical Trials Methodology Conference 2011Bristol, UK. 4-5 October 2011
In trials of physical and talking therapies or group admi-nistered treatments, clustering of patients within care-provider or treatment group has implications for samplesize and statistical analysis analogous to those found incluster randomised trials [1]. Between-cluster variationreduces precision and power when estimating treatmenteffects. Statistical analyses that fail to take account ofthis form clustering rest on the assumption of no clus-tering effect and may therefore lack conclusion validity.In trials with treatment related clustering, the cluster
size may differ systematically between treatment arms,due to differing number of therapists in each arm or dif-ferences in the size of therapy groups between arms. Theintra-cluster correlation due to therapist or therapygroup may also differ between treatment arms. Anextreme case of such heterogeneity is the partially nesteddesign in which the clustering effect is absent in onetreatment arm. Where both cluster size and the underly-ing variance components differ between treatments, fail-ure to correctly model heterogeneity can bias estimatesof the intra-cluster correlation coefficient and test size[1]. This contrasts with cluster-randomised trials wherethe distribution of cluster sizes in each treatment armwill be similar due to randomisation making cluster ran-domised trials robust to heteroscedasticity.In a cluster randomised trial, it is generally assumed
that each subject belongs to just a single unit of rando-misation so that the design is hierarchical. In non-phar-macological therapy trials patients may receivetreatment from more than one therapist or in morethan one therapy group so the multilevel model is nolonger strictly hierarchical. Statistical analysis will
therefore require simplifying assumptions regarding thepattern of clustering such as use of a primary therapistor primary group for each patient or the application ofa multiple membership model [2].In conclusion, analysis of trials with treatment related
clustering may therefore require more complex methodsof analysis than cluster randomised trials.
Published: 13 December 2011
References1. Roberts C, Roberts SA: Design and analysis of clinical trials with clustering
effects due to treatment. Clinical Trials 2005, 2:152-162.2. Browne W, Goldstein H, Rasbash J: Multiple membership multiple
classification (MMMC) models. Statistical Modelling 2001, 1:103-124.
doi:10.1186/1745-6215-12-S1-A21Cite this article as: Roberts: Statistical issues in the analysis of non-pharmacological therapy trials with clustering by care-provider ortherapy group. Trials 2011 12(Suppl 1):A21.
Submit your next manuscript to BioMed Centraland take full advantage of:
• Convenient online submission
• Thorough peer review
• No space constraints or color figure charges
• Immediate publication on acceptance
• Inclusion in PubMed, CAS, Scopus and Google Scholar
• Research which is freely available for redistribution
Submit your manuscript at www.biomedcentral.com/submitUniversity of Manchester, Manchester, M13 9PL, UK
Roberts Trials 2011, 12(Suppl 1):A21http://www.trialsjournal.com/content/12/S1/A21 TRIALS
© 2011 Roberts; licensee BioMed Central Ltd. This is an open access article distributed under the terms of the Creative CommonsAttribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction inany medium, provided the original work is properly cited.
