TL;DR
This paper introduces a novel statistical framework, Tukey's representation, for handling "non-ignorable" missing data, where the probability of data being missing depends on the missing values themselves. This approach offers a more realistic and flexible way to model such data compared to existing methods, particularly in complex biological datasets.
What Problem Does This Paper Address?
In many scientific fields, datasets often have missing values. A common assumption is that data are "missing at random" (MAR), meaning the probability of a value being missing depends only on the observed data. However, in reality, data can be "missing not at random" (MNAR), where the probability of missingness depends on the unobserved values themselves. This MNAR scenario is notoriously difficult to handle because assumptions about the missingness mechanism are often untestable and can lead to biased results. Traditional methods for MNAR data, like selection models and pattern-mixture models, require strong, often untestable, assumptions about either the data-generating process or the missingness mechanism itself. This paper aims to provide a more flexible and robust framework for analyzing data with MNAR mechanisms.
Methods
The core of the paper is the development and application of Tukey's representation. This approach reframes the problem by directly modeling the joint distribution of observed data and missingness indicators using conditional distributions. Instead of making strong assumptions about the entire data-generating process, Tukey's representation focuses on:
- Modeling the observed data density: This is often more straightforward and can be informed by existing knowledge.
- Modeling the missingness mechanism (selection function): This allows for more direct incorporation of expert knowledge or assumptions about why data might be missing.
The authors specifically apply this framework to exponential family models, which are common in statistical modeling. They develop a computationally tractable inferential strategy using Markov Chain Monte Carlo (MCMC) methods. A key advantage highlighted is that the normalizing constant, often a computational bottleneck, can be simplified for this class of models.
Datasets
The paper uses two biological datasets for illustration and evaluation:
- Transcriptomic data: From Pelechano and Pérez-Ortín (2010), with 14% missing data.
- Proteomic data: From Ghaemmaghami et al. (2003), with 34% missing data.
These datasets are used to demonstrate the utility of Tukey's representation in analyzing complex biological measurements where missing data is prevalent.
Key Insights and Ideas
- Tukey's representation as a hybrid: It bridges the gap between selection models (focusing on the missingness mechanism) and pattern-mixture models (focusing on the data distribution), offering a more balanced approach.
- Testable or knowledge-driven assumptions: The assumptions made within Tukey's representation are either testable from the data or can be directly informed by substantive knowledge, making the modeling process more transparent and realistic.
- Flexibility in modeling non-monotonic missingness: The framework can naturally accommodate missingness mechanisms that are not simply monotonic functions of the data, which is common in biological data (e.g., low and high values being less likely to be observed).
- Computational tractability: For exponential family models, the approach leads to a computationally feasible inferential strategy.
- Sensitivity analysis: The framework facilitates sensitivity analyses by allowing researchers to explore different plausible assumptions about the missingness mechanism.
Key Results
The paper demonstrates the effectiveness of Tukey's representation through simulations and real-world biological data analysis:
- Accurate recovery of parameters: In simulations, Tukey's representation accurately recovers posterior distributions for estimands, even when the true data-generating process is complex.
- Robustness to misspecification: The approach shows robustness when the assumed model for the observed data is an approximation of the true underlying distribution.
- Biological data analysis: When applied to transcriptomic and proteomic data, Tukey's representation provides estimates that are consistent with existing findings and offers a more flexible analysis compared to traditional selection models, especially in capturing complex relationships like translational regulation. The analysis of dynamic range ratios in protein and mRNA levels showed results consistent with amplification of protein levels.
Limitations and Future Work
- Positivity Condition: Tukey's representation requires that the support of the missing data is contained within the support of the observed data. Cases where this condition is violated (e.g., truncated normal models) cannot be directly estimated.
- Expert Knowledge Elicitation: While Tukey's representation simplifies prior elicitation compared to other methods, translating expert knowledge into functional forms for the missingness mechanism can still be challenging. The authors suggest using strategies for expert prior elicitation in conjunction with their method.
- Generalization: The paper focuses heavily on exponential family models. The authors conjecture that similar relationships might hold for models outside of exponential families, suggesting future work in this direction.
- Multiphase Inference: The authors suggest that Tukey's representation could be useful in multiphase inference, particularly when preprocessing data where the missingness mechanism is better understood than the underlying scientific model.
Summary updated