Standard treatment of locally advanced rectal cancer has long been based on neoadjuvant radiochemotherapy (RTCT), followed by surgery and then adjuvant chemotherapy (AC). Since the PRODIGE 23 trial, neo-AC prior to RTCT has been increasingly used, calling into question the interest of post-resection AC. To assess its effect on disease-free survival in this new context, a combined analysis of data from two clinical trials was implemented. Certain methodological challenges have been overcome, in particular the hierarchisation of data due to the fact that individuals belonged to one of the two studies. Our work compared methods accounting for trial of origin variability in pooled data.

We conducted a retrospective analysis using data from the GRECCAR 4 and PRODIGE 23 randomized clinical trials. 489 patients were included, 98 patients from the first trial and 391 from the second. Patients’ trial of origin was accounted in three approaches: (1) as a confounding factor, integrated into the calculation of the propensity score (PS), which was then used to adjust a Cox model; (2) as a random effect, modeling between-trial variability and accounting for the hierarchical structure of the data, using a PS-adjusted mixed Cox model and (3) finally as a stratification factor, with separate analyses for each trial, followed by a combination of results through a meta-analysis approach. These three methods were compared to each other in terms of result consistency and estimation precision.

Regardless of the method used, the findings were much the same. With the first method (1), the risk of disease occurrence was minimized but not significantly with the use of AC (HR 0.68; 95% CI 0.46–1.01; p=0.05). Consistent findings were noted in the second (2) (HR 0.74; 95% CI 0.50–1.09; p=0.13) and the third method (3) (HR 0.70; 95% CI 0.48–1.03; p=0.07). Moreover, the second method (2) showed a prediction interval for the variability between trials of [0.45–1.83], reflecting heterogeneity in the treatment effect across trials, as did the third method (3), revealing a beneficial effect of AC in PRODIGE 23 (HR 0.61; 95% CI 0.40–0.92; p=0.01), and a harmful effect, though non-significant, in GRECCAR 4 (HR 1.92; 95% CI 0.64–5.75; p=0.43).

The three methods used yielded similar and consistent results, suggesting a non-significant reduction in the risk of recurrence in patients who received AC. However, although the average effects estimated by these methods were similar, the use of multiple approaches revealed heterogeneity between trials, indicating that the effect of AC may vary depending on the context. This suggests the need for a thorough analysis of the sources of variability like the population characteristics or the disease stage. These results highlight the complexity of interpreting combined data and emphasize the importance of adopting appropriate methodological approaches to fully capture the information contained within the data. A mathematical comparison of the methods is currently underway.