Skip to main content

Archived Comments for: Adjuvant trastuzumab in the treatment of her-2-positive early breast cancer: a meta-analysis of published randomized trials

Back to article

  1. Choice of effect measure – was it appropriate?

    Sandro José Martins, Hospital Santa Izabel, Salvador - Brazil

    10 December 2007

    The analytic method and the specific formulas used in meta-analysis are affected by the choice of effect measures. Main outcome measures in adjuvant trials are the ratio of cumulative incidence rates (e.g. recurrence, death), and are statistically modeled as time-to-event data (risk ratio and hazard ratio). In this article, authors have chosen to handle such measures as plain odds ratios, using the Peto method. This implies that it was available data to complete 2 x 2 table of outcome by treatment, and confounding was negligible. Unfortunately, this may not be true.

    Let us examine consequences of author's choice in the main outcome measure (DFS) of the largest trial (HERA). In Figure 3, we are informed that there were 347 recurrences in the HERA trial. It was assumed that remaining patients (n=3,040) could not ever experience such event, resulting in OR = 0.54 (CI 95% 0.43-0.68), a benefit for trastuzumab therapy 18% larger then estimated by HERA investigators (hazard ratio = 0.64, CI 95%: 0.54-0.76). This assumption holds no parallel with intent-to-treat analysis (cf. Ellenberg JH, Drug Inform J 1996;.30:535–44).

    Absence of confounding is a weak assumption, as noticed in Table 1. There was considerable heterogeneity for nodal status and hormonal receptors in those data sets. This may also represent differing conditions generally unknown to the analyst, and completely beyond his control. It is absolutely necessary therefore to investigate heterogeneity before applying the simples of tests on the total frequencies. The effect of wrongly applying the test would invariably be to exaggerate the importance of the effect being investigated, as showed here, since heterogeneity inflates the effective error and thus diminishes the perception of the effect.

    I suspect that more realistic summary effect measures could be derived here from general variance-based methods (cf. Greenland, Epidemiol Rev 1987; 9:1-30), using available adjusted outcome measures and the corresponding 95% confidence interval for the adjusted measure.

    Competing interests

    None declared