TL;DR
This paper breathes new life into "Tukey’s representation," a forgotten way to model data that is "missing not at random" (MNAR). By factorizing the joint distribution of data and missingness indicators based on what we actually observe, the authors provide a framework where modeling assumptions are either testable or easily informed by scientific domain knowledge, making it a robust alternative to standard selection and pattern-mixture models.
What Problem Does This Paper Address?
In statistics, we often deal with missing data. If data is "missing at random" (MAR), we can usually ignore the reason why it’s missing. But in many fields—like proteomics or social surveys—data is "missing not at random" (MNAR). For example, a protein might be missing from a dataset simply because its concentration was too low for a sensor to detect.
The standard ways to handle this (Selection Models and Pattern-Mixture Models) have a major flaw: they require untestable assumptions. You either have to guess the distribution of the "complete" data (including the parts you never saw) or guess the distribution of the missing parts specifically. If your guess is wrong, your entire analysis is biased. The authors tackle the challenge of creating a model that relies on the observed data distribution, which we can actually see and verify.
Methods
The authors utilize Tukey’s representation, a non-standard factorization of the joint distribution of data ($Y$) and missingness indicators ($R$):
$$P(Y, R | \theta) \propto f(Y | R = 1, \theta_{Y|R=1}) \cdot \frac{f(R | Y, \theta_{R|Y})}{f(R = 1 | Y, \theta_{R|Y})}$$
Key Innovations:
- Observed-First Modeling: Unlike standard models, this starts with $f(Y | R=1)$, the density of the data we actually have. This allows researchers to use standard goodness-of-fit tests to ensure their model actually matches the reality of the collected data.
- Exponential-Logistic Class: The authors show that if the observed data follows an exponential family distribution (like Normal, Poisson, etc.) and the missingness mechanism is logistic (linear in the sufficient statistics), the math becomes incredibly elegant.
- Tractable Normalizing Constant: They derive a way to calculate the population fraction of observed data ($Q$), which allows for efficient Bayesian inference using Markov Chain Monte Carlo (MCMC) methods.
Datasets
- Simulated Semicontinuous Data: A complex mixture of continuous (Normal) and discrete point masses used to demonstrate that Tukey’s representation can accurately recover the true mean and standard deviation of a population even with 50% missing data.
- Yeast Biological Data:
- Transcriptomic (mRNA): Data from Pelechano and Pérez-Ortín (2010) with 14% missingness.
- Proteomic (Protein): Data from Ghaemmaghami et al. (2003) with 34% missingness.
- These datasets are notorious for MNAR issues because low-abundance molecules are harder to detect.
Key Insights and Ideas
- The "Aha" Moment for Exponential Families: The authors prove that if your observed data is in the exponential family and your missingness logic is "exponentially tilted," the missing data belongs to that same exponential family. This simplifies imputation immensely.
- Testability: Because the model is built on the observed data distribution, you can actually check if your model is wrong. In the yeast application, they showed that standard "Selection Models" failed to fit the observed data, while Tukey’s representation fit it perfectly.
- Propagating Uncertainty: In MNAR problems, you can never be 100% sure about the missingness mechanism. Tukey’s representation allows researchers to put a prior on the missingness parameters, ensuring that this fundamental uncertainty is reflected in the final results rather than hidden behind a "best guess."
Key Results
- Accuracy in Simulation: Tukey’s representation accurately recovered posterior distributions for population parameters as sample sizes increased, whereas misspecified standard models remained biased.
- Biological Discovery: When applied to yeast data, the model estimated a dynamic range ratio of 1.48. This supports the theory that translational regulation amplifies protein levels relative to mRNA levels.
- Model Fit: In the proteomic dataset, the authors demonstrated that the standard Selection Factorization (assuming log-normality) resulted in a visible "lack-of-fit" to the observed data, while their approach remained flexible enough to capture the true distribution.
Limitations and Future Work
- The Positivity Condition: Tukey’s representation requires that the support of the missing data is contained within the support of the observed data. It cannot handle "hard truncation" (e.g., if it is physically impossible to see any value below 5.0).
- Tail Sensitivity: The model is sensitive to how well you estimate the "tails" of the observed distribution. If the odds of missingness increase rapidly in the tails, small errors in the observed model can lead to unstable estimates.
- Future Directions: The authors suggest extending this to non-exponential family models and exploring "multiphase inference," where preprocessing steps (like data cleaning) might be better handled using this conditional framework.
Summary updated