The dose-response relationship is considered as one of the major criteria when the causality of a relationship between an exposure and an outcome is discussed. This criterion is mainly studied in observational studies but may be also of interest in Randomized Controlled Trial (RCT), particularly if the administered dose of treatment varies over time and between patients. Maintenance trials with a long follow-up meet these characteristics. We reanalyzed a maintenance trial in digestive oncology (PRODIGE 9 trial) in order to explore the dose-response relationship between the maintenance treatment and overall survival (OS).

The PRODIGE 9 study was a maintenance trial assessing the benefit of maintaining bevacizumab during chemotherapy-free intervals (CFI) in patients with a stage IV colon cancer. We defined five time-updated maintenance exposures: cumulative exposure (CE); relative exposure (RE) either expressed as a continuous (patient's CE divided by the standard deviation of the distribution of CE in the population at risk) or as a categorical variable (‘with quantiles of the distribution of the CE as cut-offs); and two variables that took into account delay since the administration thus allowing a decrease exposure over time: time-updated theoretical blood concentration (TBC) and weighted cumulative exposure (WCE). We defined these exposures using either bevacizumab received during the entire follow-up (total dose) or only during CFI, resulting in ten different exposures. These variables were then included in independent Cox models with an origin fixed at the beginning of the maintenance treatment, and with overall survival (OS) as the outcome. All models were adjusted on the group of randomization (group) and compared using the Akaike Information Criterium (AIC).

Of the initial 494 patients included in the RCT, 382 were alive and had not experienced progression at the beginning of the maintenance treatment and were thus included in the present study. All exposure variables (CE, RE, TBC and WCE) improved the quality of the model that include only the group of randomization. The best AIC was obtained when adjusting on TBC and WCE. RE led to a moderate improvement of the goodness of fit and CE only led to a small improvement. WCE was associated with a strong protective effect during the first twenty days after administration, and after 200 days. For continuous and categorical RE the highest risk was observed for medium RE, in comparison with low and high REs which were associated with a lower risk. We did not observe major differences between models using the total dose of bevacizumab and those using only the dose received during CFI. Excluding the group of randomization also did not change the estimated effects.

Of the many significant results observed in this reanalysis, only the effect of RE can be considered as a possible evidence of a protective effect of the treatment on overall survival. Indeed, even if TBC and WCE led to the best goodness of fit, their definition can induce immortal time bias and can lead them to act as a mediator variable. The protective effect of low RE is difficult to explain but might be related to an early immortal-time period. Adjustment on the treatment group did not modify the dose-effect relationship. Further analysis should rely on more complex models like structural equation models to disentangle the real impact of the administered dose of bevacizumab on outcomes.