📊 statistics

Inference on the Significance of Modalities in Multimodal Generalized Linear Models

This paper proposes a novel entropy-based metric and a consistent deviance-based statistic to enable rigorous statistical inference, including p-values and confidence intervals, for assessing the significance of individual modalities within high-dimensional multimodal generalized linear models.

Original authors: Wanting Jin, Guorong Wu, Quefeng Li

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wanting Jin, Guorong Wu, Quefeng Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a complex mystery, like diagnosing a patient's illness or predicting a stock market trend. To do this, you have a team of detectives, but they come in different "modalities" or groups. One group uses MRI scans (pictures of the brain's structure), another uses PET scans (pictures of brain activity), and a third might use genetic data.

In the past, statisticians were great at building a "super-detective" team that combined all these groups to make the best prediction possible. However, they struggled with a specific question: "How much did just the MRI group actually help? Was the PET group the real hero, or did the MRI group do all the heavy lifting?"

Existing tools could tell you if a single variable (like one specific pixel in an MRI) mattered, but they couldn't handle the complexity of asking if an entire group of variables (the whole MRI modality) was significant, especially when there were thousands of variables involved.

This paper introduces a new tool to answer that question. Here is the breakdown in simple terms:

1. The New Metric: "Expected Relative Entropy" (ERE)

The authors created a new way to measure "information gain." Think of it like a fuel gauge for your model.

The Full Tank: Imagine your model using all the data (MRI + PET + Genetics). This is your full tank of fuel.
The Partial Tank: Now, imagine you remove the MRI data and run the model with just PET and Genetics. This is a smaller tank.
The Difference: The "Expected Relative Entropy" measures exactly how much "fuel" (information) you lost when you took the MRI group out. If the number is high, the MRI group was crucial. If it's low, the model didn't really need them.

The paper proves that this metric is mathematically sound and connects to older, familiar ways of measuring model quality (like $R^2$ in simple statistics), but it works for these complex, multi-group scenarios.

2. The Challenge: Too Many Variables

The problem is that in modern science, each "detective group" (modality) can have thousands of variables. It's like having a library with millions of books, but only a few contain the answer.

The Problem: If you try to test if the "MRI Library" is important by looking at every single book, the math breaks down.
The Solution (The Two-Step Filter): The authors propose a clever two-step process to handle this:
1. The Big Sweep (Screening): First, they use a quick, rough filter (called "Sure Independence Screening") to toss out the thousands of variables that clearly don't matter. This shrinks the library down to a manageable size.
2. The Fine-Tooth Comb (Penalization): Then, they use a more precise method to fine-tune the remaining variables, ensuring they don't accidentally keep "noise" that looks like a signal.

3. The Result: Confidence Intervals and P-Values

Once they have this new measurement (the ERE), the paper provides a way to calculate a confidence interval and a p-value.

In plain English: This allows researchers to say, "We are 95% confident that the MRI data contributed this much to our model's success," or "There is a less than 1% chance that the MRI data was useless."
Why it's special: Most high-dimensional statistics require you to perfectly pick the right variables first. If you pick the wrong ones, your test fails. The authors show that their method works even if the variable selection isn't perfect. It's robust, like a car that keeps running even if you put in slightly the wrong fuel mix.

4. Real-World Test: The Alzheimer's Study

To prove their tool works, the authors applied it to real data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).

The Setup: They looked at patients with two types of brain scans: Amyloid-PET (which looks for protein clumps) and FDG-PET (which looks at how much sugar the brain cells are eating, i.e., metabolism).
The Goal: They wanted to see which scan was better at predicting three things: Memory scores, Executive function scores, and the diagnosis of Alzheimer's.
The Finding: Their new tool calculated that for predicting memory loss and diagnosis, the FDG-PET (metabolism scan) provided significantly more information than the Amyloid-PET scan.
The Takeaway: This suggests that for monitoring how a patient's thinking is declining, looking at how the brain uses energy is currently a more powerful indicator than looking at the protein clumps.

Summary

This paper gives scientists a rigorous "ruler" to measure the value of different types of data in a complex model. It solves the problem of "too many variables" by using a smart filtering system and allows researchers to confidently say, "This specific type of data is statistically significant and adds real value," without needing to perfectly identify every single important variable first.

Technical Summary: Inference on the Significance of Modalities in Multimodal Generalized Linear Models

Problem Statement
Multimodal data, which integrate information from multiple sources (e.g., MRI, PET, CT, or multi-omics), are increasingly prevalent in fields like biomedical research and artificial intelligence. While fused multimodal models often outperform single-modal models in prediction, a critical gap exists in rigorous statistical inference: there are no established tools to quantify the significance of a single modality within a high-dimensional multimodal model. Existing high-dimensional inference methods (e.g., de-biased LASSO, decorrelated score tests) are designed for testing individual covariate coefficients or their linear combinations. They cannot directly assess the contribution of an entire modality because a modality's contribution depends not only on the magnitude of its coefficients but also on complex population parameters, such as correlations between modalities. Furthermore, classical goodness-of-fit metrics (e.g., $R^2$ , pseudo- $R^2$ ) quantify the contribution of all variables collectively, not individual modalities.

Methodology
The authors propose a framework for Multimodal Generalized Linear Models (MGLM) where the response $y$ depends on a concatenated vector of variables $x = (x_1^T, \dots, x_M^T)^T$ from $M$ modalities.

Metric Definition (Expected Relative Entropy - ERE):
The authors introduce a metric, $H_m$ , to quantify the information gain of the $m$ -th modality given all others. This is defined as twice the expected Kullback-Leibler (KL) divergence between the full model (using all modalities) and a reduced "oracle" model (using all modalities except the $m$ -th):
$H_m = 2E_x [D_{KL}(p(y|x) \parallel p(y|x_{-m}))]$
where $x_{-m}$ represents the subvector of predictors excluding the $m$ -th modality. The paper demonstrates that $H_m$ is non-negative and monotone. For specific cases (e.g., linear regression), $H_m$ has a closed-form expression involving coefficients and covariance structures. For non-linear link functions (logistic, Poisson), it is defined via the expectation of the log-likelihood ratio. The authors show that conventional goodness-of-fit metrics are special cases of this ERE when all covariates are treated as a single modality.
Estimation Procedure:
To estimate $H_m$ in high-dimensional settings ( $p \gg n$ ), the authors propose a two-step deviance-based estimator:
- Step 1 (Screening): A Sure Independence Screening (SIS) procedure is applied to reduce the ultra-high dimensional space to a moderate dimension. Variables are ranked by their marginal likelihood contributions, and a threshold $\gamma_n$ is selected (via BIC minimization) to ensure the screened set contains the true support with high probability.
- Step 2 (Partial Penalization): A partially penalized maximum likelihood estimation is performed.
  - For the full model, variables in the screened set outside the $m$ -th modality are penalized (using a folded-concave penalty like SCAD), while variables within the $m$ -th modality are left unpenalized to avoid bias.
  - For the reduced model, the same procedure is used, but variables in the $m$ -th modality are forced to have zero coefficients.
    The estimator $\hat{H}_m$ is calculated as the normalized difference in deviance between the full and reduced models.
Theoretical Inference:
The authors establish the asymptotic properties of $\hat{H}_m$ :
- Consistency: $\hat{H}_m$ is consistent for $H_m$ with an error bound of $O_P(\max\{s n^{-1} \log n, n^{-1/2}\})$ .
- Asymptotic Distribution: Under regularity conditions (including sub-Gaussianity of covariates and specific rates for tuning parameters), $n\hat{H}_m$ converges to a non-central chi-squared distribution ( $\chi^2_{e_{sm}, \Gamma_n}$ ), where $e_{sm}$ is the size of the screened set for the $m$ -th modality and $\Gamma_n$ is a non-centrality parameter related to the true coefficients.
- Significance Testing: This distributional result allows for the construction of confidence intervals for $H_m$ and the calculation of p-values to test the null hypothesis $H_m = 0$ . Crucially, the authors prove that this inference does not require variable selection consistency (i.e., the estimators do not need to perfectly identify the true zero/non-zero coefficients), a significant relaxation compared to existing high-dimensional inference methods.

Key Results

Simulation Studies: The authors evaluated the method on linear, logistic, and probit regression models with $n=300$ and $p=600$ . The proposed "SIS+SCAD" method demonstrated satisfactory empirical coverage probabilities (close to the nominal 95%) for the confidence intervals of $H_m$ , particularly when the true information gain was substantial. The method outperformed a "SIS+refit" approach (which lacks the partial penalization) in terms of specificity and coverage, especially for smaller effect sizes. The results confirmed that the method maintains validity even when variable selection is imperfect.
Real Data Application: The method was applied to a multimodal neuroimaging dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI) involving 669 subjects with Amyloid-PET and FDG-PET scans. The outcomes were memory scores (MEM), executive function scores (EF), and a diagnostic label (DX).
- The analysis revealed that FDG-PET provided significantly more information than Amyloid-PET for predicting executive function and diagnostic status (lower p-values and higher ERE estimates for FDG-PET).
- Conversely, Amyloid-PET showed a stronger information gain for memory scores.
- These findings align with clinical understanding that metabolic activity (FDG) is a strong biomarker for functional decline, while amyloid burden is more specific to memory-related pathology.

Significance and Claims
The paper claims to provide the first rigorous statistical inference tools for assessing the significance of individual modalities in high-dimensional multimodal generalized linear models. By defining the Expected Relative Entropy (ERE) and deriving its asymptotic distribution, the authors enable researchers to:

Quantify the specific information gain of a modality beyond what is provided by others.
Construct confidence intervals and p-values for this gain without relying on the difficult-to-achieve condition of perfect variable selection.
Extend the concept of goodness-of-fit from the variable level to the modality level, bridging the gap between classical metrics and modern high-dimensional multimodal analysis.

The authors note that while the current work focuses on GLMs, the entropy-based metric could potentially be extended to other models (e.g., Cox proportional hazards), though such extensions are left for future research.