|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Journal of Clinical Oncology, Vol 23, No 33 (November 20), 2005: pp. 8537-8538 © 2005 American Society of Clinical Oncology. DOI: 10.1200/JCO.2005.03.6798
In Reply:Hospital Celso Pierro/PUC-Campinas, Sao Paulo, Brazil
University of Rochester School of Medicine and Dentistry, and the James P. Wilmot Cancer Center, Rochester, NY
H. Lee Moffitt Cancer Center & Research Institute, University of South Florida, Tampa, FL Dr Fridrik raises an important issue relating to the results of systematic reviews with associated meta-analyses. Clearly, the treatment effect observed in clinical trials can be expressed in several different ways. Data from therapeutic trials have generally been found to fit relative-risk models better than models of absolute risk difference. This is to say that treatments across clinical settings including oncology have been found to reduce risk most consistently by some proportion of the control risk rather than by some absolute difference. Therefore, estimates of treatment effect are most reasonably expressed as a proportional reduction in risk rather than an absolute risk difference. Therefore, the results of the meta-analysis of therapeutic use of colony-stimulating factors (CSFs) and most meta-analyses, including those of the Cochrane collaboration, are based on relative risk models.1 Often, as in this case, odds ratios are employed as estimates of relative risk for dichotomous outcomes. Sprott has demonstrated that test statistics based on the log odds ratio correspond most closely to the asymptotic normal distribution utilized for generating confidence limits in such analyses.2 Additionally, problems can arise in evaluating risk differences constrained to values from –100% to +100%, as confidence intervals will often include values outside of this range. Perhaps the most important implication for meta-analyses such as that reported here is that there is often significant heterogeneity for outcomes based on risk differences that are not seen for relative risk measures. Generally, homogeneity of treatment effect using one model indicates heterogeneity based on the other.3 Heterogeneity based on difference measures is most commonly observed when the control risk extends across a wide range of values. In the meta-analysis of therapeutic CSFs, the significant heterogeneity observed for the outcomes of interest when expressed as risk differences completely disappears or becomes much less significant when expressed as odds ratios or relative risk.
Despite the above methodological reasons for the analysis and presentation as set forth in the article, relative risk measures tell us nothing about the absolute risk and risk reduction observed. As clinicians, we share Dr Fridik’s interest in information concerning the absolute risks and risk differences as these may also influence clinical decision making. Although absolute risks and risk differences can be readily generated from Table 3 of the article, it does require some calculation. Prolonged hospitalization ( Authors’ Disclosures of Potential Conflicts of Interest
Although all authors completed the disclosure declaration, the following author or immediate family members indicated a financial interest. No conflict exists for drugs or devices used in a study if they are not being evaluated as part of the investigation. For a detailed discription of the disclosure categories, or for more information about ASCO's conflict of interest policy, please refer to the Author Disclosure Declaration and the Disclosures of Potential Conflicts of Interest section in Information for Contributors.
Dollar Amount Codes (A) < $10,000 (B) $10,000–99,000 (C)
Acknowledgment This review was undertaken under the auspices of the Cochrane Collaboration. Local University funds were utilized to support the costs of study. However, Dr Lyman receives research grant support from Amgen. REFERENCES
1. Clark OAC, Lyman GH, Castro AA, et al: Colony stimulating factors for chemotherapy induced febrile neutropenia: A meta-analysis of randomized controlled trials. J Clin Oncol 23:4198-4214, 2005
2. Sprott DA: Normal likelihoods and their relation to large sample theory of estimation. Biometrika 60:457-465, 1973 3. Whitehead A: Meta-Analysis of Controlled Clinical Trials. Hoboken, NJ, John Wiley & Sons, 2002, pp 336
Related Correspondence
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2005 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
10 days) was observed in 32% of control subjects and 23% of patients receiving CSFs for an absolute risk difference of 10% (95% CI, 1.0% to 19.0%) and a number needed to treat (NNT) to prevent prolonged hospitalization of 10 (P = .0004). Infection-related mortality was observed in 5.7% and 3.1% in the control and CSF-treated arms, respectively, for an absolute risk difference of 3.4% (95% CI, 0.0% to 6.8%) and an NNT of 33 to prevent one infection-related death (P = .05). There was no significant difference in overall mortality in these trials (7.1% v 5.1%; P = .117). Patients included in these randomized controlled trials were selected such that those individuals with major comorbidities or complications were generally excluded. These absolute risk estimates, therefore, most likely underestimate the risk in the general population of cancer patients admitted with febrile neutropenia while the relative risk reductions may well generalize across levels of risk. We also emphasize that the absolute risk estimates are associated with considerable heterogeneity and the statistical analysis presented in the paper was based on relative risk rather than absolute difference assumptions. Nevertheless, these calculations may be helpful for clinical understanding and discussion if not for formal inference or comparison. 
