How Long Should We Wait? The Impact of Immature Data on Colorectal Cancer Decision Modeling.

Objective

To identify a practical data maturity threshold at which model-based long-term survival projections become sufficiently reliable to inform health technology assessment (HTA).

Methods

This retrospective modeling study used real-world, patient-level data from the Korea Clinical Data Utilization Network for Research Excellence registry. Patients with KRAS wild-type advanced colorectal cancer receiving first-line cetuximab- or bevacizumab-based chemotherapy between 2013 and 2021 (N = 1,208) were included. Hypothetical immature datasets were created by right-censoring at 30%, 50%, and 70% maturity (30%, 50%, and 70% of deaths observed). Partitioned survival analysis (PartSA) and state-transition model (STM) were developed for each maturity level and validated against observed 6-year overall survival and life expectancy using absolute prediction error and coverage within observed 95% confidence intervals.

Results

With highly immature data at 30% maturity, 6-year survival prediction errors ranged from 2.4%-11.1% (PartSA) and 4.5%-16.3% (STM). Prediction stability improved substantially at 50% maturity (PartSA 0.4%-5.8%, STM 0.04%-7.7%), with only modest gains at 70% maturity (0.2%-3.9% and 0.3%-6.1%). Life-expectancy deviations showed a similar pattern, narrowing from up to 0.4 years (PartSA) and 0.8 years (STM) at 30% maturity to within ±0.3 years at 50% and ±0.2 years at 70% maturity, regardless of modeling framework.

Conclusions

In this first-line advanced colorectal cancer case study, prediction stability improved substantially once 50% maturity was reached. This threshold may provide a practical benchmark for planning reassessment or evidence-updating strategies in HTA; however, its applicability to other cancers, treatment settings, and data structures requires further evaluation.