We read with interest the meta-analysis of fixeddose studies on the efficacy of PDE5 inhibitors by Berner and colleagues (Int J Impot Res 2006; 18: 229-235). The authors suggested superiority of sildenafil compared to vardenafil based on relative improvements in IIEF-EF domain scores. 1 However, their analysis fails to consider the following issues.The studies had different durations of treatment. Although the modal duration was 12 weeks, three of the studies were for 3, 4 and 8 weeks and one at 26 weeks. These differences in study duration could have influenced outcomes in several ways, including increased discontinuation rates or disease progression in the longer duration studies.The sildenafil studies were conducted, on average, about 3 years earlier than the tadalafil and vardenafil studies. As a result, the patients were PDE5-inhibitor-naïve, which may have made them more responsive to treatment. For example, a recent study in PDE5-inhibitor-naïve men with ED and hypertension showed average EF domain score improvements with vardenafil of 10.7 points from baseline. 2 The authors acknowledge that 'in cases of insufficient data, standard deviations (SD) were substituted. The estimation was based either on figures presenting confidence intervals (CIs), standard errors, or SDs or on other studies with the same agent, which reported SD for treatment and placebo group.' Imputing data broadly in this manner could result in a significant degree of bias in the comparison.Several tests of heterogeneity were performed. A value of 19.47 for Cochran's Q 3 was reported with an associated value of 33.23% (assuming 13 degrees of freedom) for Higgins' I. 2,4 Although the test failed to reject the null hypothesis of homogeneity, it has been noted that this test lacks adequate power. 5 As a result, investigators often use a ¼ 0.10 as a threshold for statistical significance when performing the Q-test. At this level, the authors' test result raises questions about the appropriateness of pooling the studies. The funnel plot also suggests the presence of a moderate degree of heterogeneity across studies.A related statistical concern is their selection of a fixed effects model (FEM) for the meta-analysis. The FEM assumes that variability in the study results is exclusively the result of random variation. 6 In contrast, a random effects model (REM) 7 assumes different underlying sources of variance for each study and considers this as a potential source of variation in the model. When there is substantial between-study variation, REM will tend to yield larger CIs than FEM. If an REM model had been used, it is conceivable that observed differences among the agents would not have been statistically significant. The authors should ideally have tested both models and, in the event that different results were obtained, reported both sets of findings. This is commonly performed in meta-analyses published in top-tier journals. 8,9 The IIEF-EF domain score can be assessed in one of three ways: (1) absolute values at baseline and end point...