Treatment-By-Center Interaction: What is The Issue?
- PDF / 557,414 Bytes
- 10 Pages / 504 x 719.759 pts Page_size
- 24 Downloads / 149 Views
0092-8615/97 Copyright 0 1997 Drug Information Association Inc.
TREATMENT-BY-CENTER INTERACTION: WHAT IS THE ISSUE? ANDERSK A L L ~ N Astra Draco AB, Lund, Sweden
This paper discusses treatment-by-center interactions. It is argued that from a clinical perspective this is a nonissue and that the statistical analysis should reflect this. A mode of analysis of multicenter studies is proposed that focuses on the units studied and patients, and still acknowledges the statistician’s wish to adjust for centers and their interaction with treatment. The analysis is not conditioned on the outcome of the interaction test. Key Words: Multicenter studies; Treatment-by-center-interactions;ANOVA
INTRODUCTION WHEN ANALYZING MULTICENTER studies, one of the issues is how to handle the treatment-by-center interaction (1). This is an issue regardless of whether the interaction is significant or not. If it is not significant, should researchers analyze without the interaction term in the model or should it be included? If it is significant, what, if any, conclusions can be drawn from the study? Most of these issues will be challenged by the claim that there is no special issue involved with the fact that the study is a multicenter, as compared to a single center, study. In both cases, a heterogenous population of patients is studied. The problem will be approached from both a clinical perspective and a mathematical perspective.
approach to a parametric analysis is to define the model:
Y = Treatment + Center + TreatmentCenter. First the P-value for the interaction test is computed and the analysis proceeds as follows:
1. If P > 0.10, it is declared that there is no interaction between treatment and center and thus one can talk about a treatment effect. Some statisticians keep the original model, and compute treatment differences from this model. Other statisticians remove the interaction factor and analyze the remaining model, and 2. If P < 0.10, it is declared that there is an interaction between treatment and center and therefore one cannot talk about a treatPROBLEM STATEMENT ment effect (since it is center dependent). Consider a multicenter study in which some To remedy this, statisticians have develeffect variable Y is measured. A standard oped two sorts of interactions: quantitative and qualitative. The former means that the results go in the same direction at most Resented at the DIA Workshop ‘Statistical Methodolcenters, so that the interaction could be ogy in Clinical RBrD.” April 15-17, 1996. Copenhagen, considered to be a scale problem, that is, Denmark. if the effect variable had been measured Reprint address: Anders Killltn, Astra Draco AJ3. PO Box 34. S-221 00 Lund, Sweden. on a different scale, there might be no 92 7 Downloaded from dij.sagepub.com at The University of Auckland Library on May 17, 2015
Anders Kdll4n
928
interaction. Qualitative interaction means that the results actually differ at different centers and this is considered to be serious, so serious, in fact, that many statisticians do not want to ta
Data Loading...
