© 2007 American Thoracic Society
From the Authors:We appreciate Prof. Suissa's interest in our article (1) and would like to contrast two approaches based on the data of our clinical trial. The intended primary analysis for the exacerbation rate in the study protocol was based on a Poisson regression model adjusting for potential or probable prognostic factors such as sex, age, smoking status, and baseline FEV1 (% predicted) without inclusion of an overdispersion parameter. The model itself, however, does not assume that all patients in a treatment group come from a population characterized by a single rate of exacerbation. Furthermore, at the time that the study protocol was completed, there had been little, if any, advice with regard to the inclusion of the overdispersion parameter in the analysis models under discussion. In accordance with widely accepted methodological standards, we provided the primary analysis as determined in the study protocol to avoid any speculations of data-driven changes. Nevertheless, as recommended by Suissa (2), we also performed a post hoc reanalysis of the exacerbation rates including an adjustment for overdispersion. The estimated ratio is again 0.65 and the corresponding value of the test statistic decreases, as noted by Suissa in his letter. Interestingly, the associated P value is again less than 0.0001, consistent with the result in our article (1). As expected by Suissa, the confidence interval increases, but only slightly due to the "overdispersion" adjustment, and it is now (0.54, 0.79). Thus, the length of the confidence interval increases marginally by 0.06. Both analyses give very similar results. Obviously, in real clinical trials the claimed general underestimation of P values based on emulated trial data (2) is not in every case of considerable magnitude. When modeling existing data, researchers have to decide not only whether to include an overdispersion parameter but also, for example, whether to use negative binomial modeling or Poisson modeling (3). Thus, to date, when modeling data from real clinical trials, there is no general "true" or "false" means of analysis available to researchers.
Maingau Hospital, Frankfurt am Main, Germany
Frankfurt am Main, Germany
Hannover, Germany
Philipps University, Marburg, Germany FOOTNOTES
* M.W. and T.G. were employees of GlaxoSmithKline Germany at the time of the study. Conflict of Interest Statement: P.K. has been reimbursed by Altana, AstraZeneca (AZ), GlaxoSmithKline (GSK), Novartis, and Viatris for attending several conferences; he has participated as a speaker in scientific meetings and courses organized and financed by various pharmaceutical companies (Altana, AZ, Berlin-Chemie, Boehringer Ingelheim [BI], Cefalon, GSK, Novartis, Pfizer); and served on advisory boards for the following companies: AZ, BI, GSK, Novartis, Pfizer. He has received aggregated honoraria of $12,000 in 2005 from GSK and $11,000 from Novartis; his group has received $14,000, $18,000, and $15,000 in 2003, 2004, and 2005, respectively, from Altana, AZ, GSK, and Novartis. M.W. was an employee of GSK Germany at the time of the study. T.G. was an employee of GSK Germany at the time of the study. C.V. has given presentations in the last 3 years at industry symposia sponsored by Altana, AZ, Bayer, BI, GSK, Merck Darmstadt, Novartis, Pfizer, Sanofi Aventis, and Talecris; he also served on advisory boards for Altana, AZ, Bayer, BI, GSK, Pfizer, and Sanofi Aventis; in addition, his institution took part in clinical studies sponsored by AZ, GSK, and Intermune; in 2005, aggregated honoraria and industry-sponsored grants comprised $22,000 from Altana, $50,000 from AZ, and $30,000 from GSK. REFERENCES
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
