Interpreting Interaction: The Classical Approach

  • PDF / 357,976 Bytes
  • 6 Pages / 504 x 719.759 pts Page_size
  • 41 Downloads / 229 Views

DOWNLOAD

REPORT


0092-8615198 Copyright Q 1998 Drug Information Association lnc.

Printed in the USA. All rights reserved.

INTERPRETING INTERACTION: THE CLASSICAL APPROACH STEVENM. SNAPINN Merck Research Laboratories, West Point, Pennsylvania

This paper describes a variety of issues in the design and analysis of multiclinic trials, focusing in particular on the detection and interpretation of treatment-by-clinic interaction. First Fleiss, which represents the classical approach and covers such issues as pooling the data, fired versus random effects, and the analysis of main effects in the presence of interaction, is reviewed. Next, different methods for distinguishing quantitative from qualitative interaction are reviewed, and the influence of scale transformations on the interpretation of interaction is discussed. Finally, nonparametric procedures for detecting treatment-by-center interaction are reviewed and the advantages of the h a r d Conover-type ranking procedure are described. Key Words: Multiclinic trials; Treatment-by-clinicinteraction; Qualitative interaction;

Nonparametric tests

INTRODUCTION

THIS PAPER REVIEWS and discusses the relevant statistical issues in multiclinic trials, focusing in particular on the detection and interpretation of treatment-by-clinic interaction. Many of these issues were very clearly described by Fleiss in his influential paper on the analysis of data from multiclinic trials (1). Therefore, that paper will first be reviewed. Following that two additional issues will be discussed: the distinction between quantitative and qualitative interaction and the nonparametric detection of interaction. REVIEW OF FLEISS (1) The most common reason for performing a clinical trial at multiple sites is to achieve

Presented at the DIA Workshop “Statistical Methodology in Clinical Research and Development,” April 15-

17. 1996, Copenhagen, Denmark.

Reprint address: Steven M. Snapinn, Merck Research Laboratories, BL-3,West Point, PA 19486.

the desired enrollment in a short period of time. In addition, while clinical trial populations are not necessarily very representative of the general patient population, populations from multiclinic trials tend to cover a broader spectrum than populations from a single center, and thus the results of multiclinic trials are expected to be more generalizable. The very differences in patient demographics, medical conditions, and treatment milieux (eg, certain concomitant medications or medical procedures might be more common in one region of the country than in another) that make the results of multiclinic trials more generalizable, however, also introduce the potential for the results to differ among clinics. When these differences affect the randomized treatment regimens unequally, treatment-by-clinic interaction is introduced. Other sources of treatment-by-clinic interaction include departures from the protocol and different criteria for evaluating the response at certain centers. Data from a multiclinic trial are collected separately at C clinics, and then pooled in

4