TL;DR
This paper investigates two statistical modeling approaches, mixed models and joint models, for analyzing tumor growth data in mouse clinical trials (MCTs) using patient-derived xenografts (PDXs). The authors demonstrate that joint models can produce biased results when mouse deaths (dropouts) are predictable by observed tumor volumes (MAR), a common scenario in MCTs. They conclude that mixed models are more robust and recommend their use for analyzing MCT data.
What Problem Does This Paper Address?
In preclinical cancer research, patient-derived xenograft (PDX) models are crucial for evaluating the efficacy of new drugs. Mouse clinical trials (MCTs) extend this by pooling data from multiple PDX models, mimicking the complexity of human clinical trials. A key challenge in analyzing MCT data is handling the "dropout" of mice due to death before the study's planned end. This dropout can introduce "survivor bias," where the remaining mice might not be representative of the original group, potentially skewing treatment effect estimates.
Two main statistical approaches are used to address this:
- Mixed-effects models: These models assume that the reason a mouse drops out is related to the observed tumor volume data (Missing At Random, MAR).
- Joint models: These models can account for situations where dropout is related to unobserved tumor volume data (Missing Not At Random, MNAR).
The controversy lies in which model is more appropriate for MCTs, especially when many mice die before the study concludes. This paper aims to clarify this by comparing these two methods using both real MCT data and simulations.
Methods
The study employed two primary statistical modeling strategies:
- Linear Mixed-Effects Regression Models: These models are designed to handle longitudinal data (repeated measurements over time) and account for correlations within subjects (e.g., multiple measurements from the same mouse). The authors used a piecewise linear model to capture different tumor growth rates before and after treatment, incorporating random effects at the mouse and PDX levels to account for hierarchical data structures. This approach is considered robust to MAR data.
- Joint Shared Random-Effects Models: These models combine a longitudinal model (like the mixed model) with a survival model (to predict the time to an event, such as death). They are specifically designed to handle MNAR data by linking the longitudinal marker (tumor volume) to the risk of the event. The authors used a proportional hazards model for the survival component.
To further investigate the impact of missing data mechanisms, the researchers performed a complementary analysis. They artificially modified the real MCT dataset to simulate an MNAR dropout scenario by making tumor volumes leading to sacrifice unobservable, while keeping the death dates unchanged. This allowed them to compare how the joint and mixed models performed under different assumptions about the missing data.
A simulation study was also conducted to systematically evaluate the performance of both models across various proportions of MAR and MNAR deaths, with different sample sizes.
Datasets
The primary dataset used was from a real Mouse Clinical Trial (MCT) involving 25 PDX breast cancer models and a total of 225 mice. The mice were randomized into treatment arms (vehicle, 5 mg/kg, or 20 mg/kg of liposomal irinotecan) and stratified by tumor subtypes (HR and TNBC). Tumor volumes were measured approximately every 3 days.
The paper also describes a simulation study where datasets were generated with varying proportions of MAR and MNAR deaths, and different sample sizes (N=300, 200, 100).
Key Insights and Ideas
- The "Why" of Death Matters: The critical distinction between MAR and MNAR dropout mechanisms is not just theoretical; it has a tangible impact on the estimated treatment effects.
- Sacrifices Skew Results: In MCTs, a significant portion of mouse deaths are often due to planned sacrifices when tumors reach a predefined ethical size. This type of sacrifice is a classic example of MAR data, as the decision to sacrifice is based on observed tumor volume.
- Joint Models Can Be Misled by MAR: When a joint model is applied to data where deaths are primarily MAR (like sacrifices), it can incorrectly attribute the dropout to an MNAR process, leading to biased estimates of treatment efficacy.
- Mixed Models Offer Robustness: Mixed models, which are designed for MAR data, appear to be more reliable when the dropout mechanism is indeed MAR, providing more stable and less biased estimates.
- Complexity vs. Robustness: While joint models offer a more sophisticated framework for handling complex missing data scenarios, their complexity can also lead to convergence issues and a higher risk of misspecification, especially with limited data.
Key Results
- In the real MCT dataset, the majority of deaths (88%) were sacrifices due to reaching a large tumor size, indicating a MAR mechanism.
- The joint model estimated a significantly different tumor growth slope (-0.02 log(mm³)/week) compared to the mixed model (0.24 log(mm³)/week) when analyzing the original data, highlighting the impact of the MAR dropout.
- When the data was artificially perturbed to simulate an MNAR scenario, the estimates from the joint and mixed models became much closer, supporting the idea that the bias in the joint model was due to the MAR nature of the original data.
- The simulation study showed that mixed models converged more often than joint models, especially when there was a mix of MAR and MNAR deaths. Joint model slope estimates were more biased in these mixed scenarios.
Limitations and Future Work
- Model Complexity and Convergence: The authors note that joint models are complex and can have convergence issues, particularly with smaller datasets. While they explored ways to improve convergence in simulations, real-world application might still face challenges.
- Parametrization of Joint Models: Joint models can be parameterized in various ways, increasing the risk of misspecification and subsequent biased estimates.
- Future Work: The authors suggest that future research could explore more complex tumor growth dynamics and treatment effects using mechanistic models, and investigate how the choice of statistical model impacts these more intricate analyses. They also emphasize the importance of carefully recording the reasons for dropout to inform the choice of statistical model.
Summary updated