Response

We thank Freidlin and Korn for their comments about our article exploring the impact of a personalized rationale upon treatment outcomes in cancer drugs approved by the UF Food and Drug Administration (FDA) (1). The first issue raised is regarding the interpretation of the findings because of the exclusion of non-FDA-approved agents. In fact, we acknowledge this issue as a limitation for the generalizability of the results in our discussion section. As in any meta-analysis, there are studies excluded based on certain criteria, which does not change or invalidate the results for the group of trials selected for inclusion. Nonetheless, we should mention that we recently published another meta-analysis exploring outcomes on phase II trials (which included approved and non-approved agents), and the results also indicated favorable outcomes with a personalized strategy (2).

A second issue that was raised is a statistical bias related to the sample size of the studies, possibly confounding the results of the overall survival analysis. Indeed, a bias estimate on hazard ratios can be a result of different sample sizes between the personalized and nonpersonalized trials (Figure 1 of the Correspondence provides an intuitive explanation). Nonetheless, we didn’t claim that the difference in survival between randomized personalized and nonpersonalized randomized trials was significant (we described the difference as non–statistically significant, P = .07). We agree that this analysis has some intrinsic biases that are hard to quantify. Some of them are also against personalized trials, including a higher crossover rate allowed for these studies, which can underestimate the benefits in survival. It is also important to note that our analyses were not restricted to the survival data in randomized trials. In order to maximize the evidence and a better estimation of the truth, we evaluated different efficacy outcomes, including progression-free survival, response rate, and relative response rate through more than one statistical approach (random effects meta-analysis and pooled analysis of time-to-event endpoints). For all of these parameters analyzed (except survival in randomized trials), a significant difference favoring personalized trials was reported, which sustained our main conclusion that personalized drugs were associated with improved efficacy outcomes amongst FDA-approved agents.

As discussed in our paper, 55 of the 112 trials leading to approval (49%) were nonrandomized. Hence, an analysis restricted to the hazard ratios of randomized trials might also have bias because some very active drugs were approved without a randomized trial (including important personalized therapies). In the median survival analysis, more trials were included compared with the random effects meta-analysis (60 vs 46 trials), and a longer median survival was reported in favor of personalized therapies (19.3 months vs 13.5 months, P = .01).

Every statistical approach, including randomized clinical trials, represents an estimation of the truth. We clearly recognize that some of our analyses have intrinsic biases that are complex. We cannot completely eliminate biases, but we believe that the consistency of the evidence in showing that personalized trials yield better efficacy outcomes is robust.

Razelle Kurzrock receives research funds from Sequenom, Guardant, Foundation Medicine, Genentech, Pfizer, and Merck Serono; receives consultant fees from Sequenom; and has an ownership interest in RScueRX.