OBJECTIVES In oncology clinical trials, the analysis of PRO data can be challenging due to informati ve and administratively censored data. A joint latent-variable and pattern-mixture modeling (PMM) approach was applied to data from an oncology clinical trial to identify the effect of censoring on the interpretation of PRO (specifically fatigue) and overall survival.
METHODS: Data came from a phase 3 clinical trial investigating gemcitabine + cisplatin versus etopside + cisplatin in the treatment of locally advanced or metastatic non-small cell lung cancer (N = 131). Fatigue was measured using the EROTC QLQ-C30. A growth-mixture modelling (GMM) approach combined with a PMM approach was used. GMM searches through the patient-level data to identify subgroups (latent classes) based on the similarity of fatigue start ing points and changes over time. This was extended with a PMM (which creates categorical variables reflecting every pattern of missing data) to identify these latent classes based also on missing data patterns. This allowed an explicit analysis of missing data and its effect on fatigue across time. Survival analysis was then con- ducted to compare the differential effect on overall survival (OS).
RESULTS: Three subgroups were identified: 1) Low Fatigue (44% of trial sample) began the trial with low levels of fatigue and showed no change over time; 2) High Fatigue (38%) began the trial with high levels of fatigue that slightly increased over time; 3) Increasing fatigue (18%) began the trial with low levels of fatigue that dramatically increased (worsened) over time. Significant differences between the emergent subgroups were found in OS: Low Fatigue had greater OS than High Fatigue and Increasing Fatigue.
CONCLUSIONS: This is the first known example that directly demonstrates how PRO response is associated with OS through use of joint latent-variable PMM. Moreover, results show that use of PMM with OS analyses provides a greater understanding of treatment effect.
