The TIVO-1 study compared tivozanib (tivo) and sorafenib (sora) in patients with renal cell carcinoma. It demonstrated statistically significant PFS benefit for tivo over sora, but a trend towards poorer OS with tivo. The authors attributed this to a crossover effect, since most sora patients crossed over to tivo on progression. We initially carried out a crossover analysis on the intention to treat dataset to assess whether this was the case. However, limitations of the techniques used meant that any sequential benefit of second line targeted therapy could not be taken into account. Further analysis was performed to explore OS benefit in post-randomized subgroups of patients who did not crossover to targeted therapy.
Firstly, Cox proportional regression from observed data for patients who did not crossover was used to derive OS benefit. Propensity score matching was then used to minimize potential bias induced by post-randomization subgroup analyses. Matching was used to create a data set closer to one resulting from a perfectly blocked (and possibly randomized) study. The exact match was used to match each patient in the tivo group to possible patients in the sora group with exactly the same values on all covariates. This data set was used to derive a Cox proportional regression of tivo vs. sora. As a sensitivity analysis, propensity score using optimal matching was performed.
Survival curves from the observed data of patients who did not crossover (211 tivo patients, 94 sora patients) set yielded a numerically better OS with tivo vs sora (HR = 0.9405, p = 0.74). The exact matching data set yielded an even more improved OS benefit with tivo vs sora, although still not statistically significant (HR = 0.84, p= 0.62). The HR for OS using the optimal matching approach was 0.99, p = 0.98.
After matching treatment groups, the HR of tivo vs. sora improved, suggesting that tivo yields a better OS benefit when data is adjusted for potential confounding variables. Although there is a discrepancy in the HR estimate between the two propensity score methods used in our analysis (exact and optimal), we suspect that the result from the exact matching is plausible given the similarity of the baseline covariates between the two groups.
Clinical trial identification
Legal entity responsible for the study
Dr Jonathan Belsey
AVEO Pharmaceuticals and Eusa Pharma
J. Belsey, E. Kemadjou: JB Medical Limited was funded by Eusa Pharma to carry out the data analysis and to produce a report of their findings.