📊 statistics

Box Confidence Depth: simulation-based inference with hyper-rectangles

This paper introduces a novel simulation-based method called Box Confidence Depth that constructs accurate multivariate confidence regions for parametric and generative models, particularly in scenarios with limited data where traditional asymptotic approximations fail, by utilizing random hyper-rectangles and a probabilistic acceptance rule to derive depth-based confidence distributions.

Original authors: Elena Bortolato, Laura Ventura

Published 2026-01-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elena Bortolato, Laura Ventura

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 a detective trying to solve a mystery, but you don't have a clear rulebook (a mathematical formula) to tell you exactly how the crime happened. You only have a vague idea of the "rules of the game" (a computer model) and a single piece of evidence found at the scene (your observed data).

Your goal is to figure out: "What are the most likely settings for the game's rules that could have produced this specific piece of evidence?"

This is the problem the paper "Box Confidence Depth" tries to solve. Here is how their new method works, explained simply.

The Old Way: Measuring the Distance

Usually, when researchers try to guess the right settings, they play a game of "Hot or Cold."

They guess a setting for the rules.
They run the computer model to see what data it produces.
They measure the distance between the model's fake data and the real evidence.
If the distance is small enough, they keep the guess. If it's too far, they throw it away.

The Problem: In complex, multi-dimensional worlds (where you have many different clues at once), measuring "distance" becomes a nightmare. It's like trying to find a specific spot in a giant, dark 3D maze by only knowing how far you are from the center. As the maze gets bigger, it becomes almost impossible to find the right spot by just measuring distance. This is called the "curse of dimensionality."

The New Way: The "Box" Method

The authors propose a clever new way to play the game. Instead of measuring distance, they use boxes (hyper-rectangles).

Here is the step-by-step analogy:

The Setup: Imagine you have a giant, invisible box floating in a room. This box represents a range of possible outcomes for your clues.
The Test: You pick a random guess for the game's rules. You run the model twice with that same guess.
- Run #1 gives you a set of clues (let's call them the "Left Edge" of a box).
- Run #2 gives you another set of clues (the "Right Edge" of the box).
- Together, these two runs define a 3D Box (or a multi-dimensional box) in the space of all possible clues.
The Decision: Now, look at your real evidence (the data you actually found).
- Is the real evidence sitting inside the box created by your two fake runs?
- YES: Great! Your guess for the rules is "plausible." We keep it.
- NO: The real evidence is outside the box. Your guess is likely wrong. We discard it.

Why "Boxes" are Better than "Distances"

Think of it like this:

Distance (Old Way): You are trying to hit a bullseye with a dart. If you miss by even a tiny bit, you fail. In high dimensions, the "bullseye" becomes so small it's nearly invisible.
Boxes (New Way): You aren't trying to hit a single point. You are checking if the real evidence falls between two other points. It's like checking if a book fits on a shelf. Even if the shelf is huge, as long as the book is somewhere between the left and right edges, it counts.

This method doesn't care about the exact "distance" between the fake data and real data; it only cares about the order. Is the real data "in the middle" of the fake data?

The Result: A Map of Confidence

After running this test thousands of times with different guesses, the computer builds a "map" (called the Box-Confidence Depth).

High Depth: Areas on the map where the real evidence frequently fell inside the boxes. These are the most likely settings for your rules.
Low Depth: Areas where the real evidence rarely fell inside the boxes. These are unlikely settings.

From this map, researchers can draw a "confidence region"—a safe zone that says, "We are 95% sure the true rules are somewhere inside this shape."

Key Features Mentioned in the Paper

No Need for a "Summary Score": Often, researchers try to squeeze all their complex data into one single number (like an average) to make the math easier. This method allows you to use all the raw data directly, which is like using the whole puzzle instead of just one piece.
Works with Many Variables: It handles situations where you have many different types of clues (multivariate) without getting confused, because the "box" logic works naturally in multiple dimensions.
The "S" Trick: To make the method faster and more accurate in very complex situations, the authors suggest using more than two fake runs (S runs) to define the box. Instead of a box defined by a left and right edge, you define it by the minimum and maximum of many runs. This creates a "fuzzier," larger box that is easier to catch the real data in, making the search more efficient.

In a Nutshell

The paper introduces a new statistical tool that replaces the difficult task of measuring "how far off" a guess is, with the simpler task of checking "is the real data inside the range of our fake data?" This creates a robust, flexible way to find the most likely settings for complex computer models, even when we don't have perfect mathematical formulas to guide us.

Technical Summary: Box Confidence Depth

1. Problem Statement

The paper addresses the challenge of constructing calibrated confidence regions for parameters in complex statistical models, particularly where the likelihood function is computationally intractable or prohibitively expensive to calculate. This scenario is common in generative models and situations with limited data where conventional asymptotic approximations fail.

Existing likelihood-free inference (LFI) methods, such as Approximate Bayesian Computation (ABC), often rely on distance metrics or divergence functions to compare observed and simulated data. These approaches face significant difficulties in high-dimensional settings due to the "curse of dimensionality," where the probability of simulated data falling within a fixed tolerance radius of observed data decays super-exponentially. Furthermore, standard frequentist approaches often require reducing multivariate data to a scalar summary statistic, potentially losing information, or rely on assumptions of stochastic ordering that may not hold in complex, non-regular settings.

2. Methodology: Box Confidence Depth (Box-CD)

The authors propose a novel simulation-based inference method called Box Confidence Depth (Box-CD). This approach constructs confidence regions without requiring a likelihood function or a scalar summary statistic, operating directly on multivariate data or statistics.

Core Mechanism

The method utilizes a rejection-sampling framework based on the concept of data depth and hyper-rectangles (boxes):

Proposal Generation: A parameter value $\theta^*$ is drawn from a proposal distribution $\pi(\theta)$ (typically uniform over a bounded subset of the parameter space).
Simulation: For each $\theta^*$ , two independent pseudo-samples, $y^*_1$ and $y^*_2$ , are generated from the model $p(y|\theta^*)$ .
Statistic Computation: Summary statistics $t^*_1$ and $t^*_2$ are computed from the pseudo-samples.
Box Construction: A $d$ -dimensional hyper-rectangle (Box), denoted $B^*_t$ , is defined by the order statistics of $t^*_1$ and $t^*_2$ along each coordinate. Specifically, for dimension $j$ , the interval is $[\min(t^*_{1,j}, t^*_{2,j}), \max(t^*_{1,j}, t^*_{2,j})]$ .
Acceptance Rule: The proposal $\theta^*$ is accepted if the observed summary statistic $t_{obs}$ falls strictly within this hyper-rectangle (i.e., $t^*_{(1),j} < t_{obs, j} < t^*_{(2),j}$ for all $j=1, \dots, d$ ).

The Box-Confidence Depth Function

The collection of accepted parameters forms an empirical distribution known as the Box-Confidence Depth, $CD_{box}(\theta)$ . This function assigns a "depth" or centrality measure to each parameter value. Higher values indicate that the parameter is more central or representative of the observed data, as it frequently generates boxes that contain the observation.

Theoretical Properties

Scalar Case: For a scalar parameter, the acceptance probability is proportional to $F_t(t_{obs}|\theta)[1 - F_t(t_{obs}|\theta)]$ , where $F_t$ is the cumulative distribution function of the summary statistic. The maximizer of this function provides a median-unbiased point estimator.
Multivariate Case: The method extends naturally to multivariate parameters and statistics. Unlike ABC, it does not require a distance metric satisfying the identity of indiscernibles. Instead, it relies on a "center-outward" ordering induced by the probability of the observed data falling within the random boxes.
Confidence Sets: A confidence set with level $1-\alpha$ is defined as the region where $CD_{box}(\theta) \geq Q_{CD_{box}}(\alpha)$ , where $Q$ is the quantile function. Theorem 3.5 establishes that these regions have exact calibration properties.
Efficiency: Under Monotone Likelihood Ratio (MLR) conditions, the Box-CD test is shown to be at least as powerful as the Likelihood Ratio Test (LRT), with the potential to be strictly more powerful when aggregating information from dependent multivariate statistics.

High-Dimensional Adaptation

To mitigate the curse of dimensionality, the authors propose an extension using $S$ pseudo-samples (where $S$ is an even number) instead of just two. The box boundaries are then defined by the minimum and maximum order statistics across the $S$ samples. While the computational cost per iteration increases, the acceptance probability decays at a slower rate ( $\propto (1/4)^d$ ) compared to the super-exponential decay of fixed-radius ABC balls, allowing for more efficient estimation in higher dimensions.

3. Key Results

The authors validate the method through simulation studies across various models, comparing the empirical coverage of Box-CD confidence sets against nominal levels and, where available, Likelihood Ratio (LR) tests.

Logistic Regression: The method achieved coverage levels closer to nominal values (e.g., 0.95) than the LR test, with significantly faster computation times (1.80s for 10,000 replications).
Multivariate t-Distribution: The method successfully maintained nominal coverage for multivariate parameters with dependent statistics, a scenario where standard reductions might fail.
Normal Mixture Model: The study demonstrated that increasing the number of pseudo-samples ( $S$ ) improves the stability of the Box-CD function and increases the number of accepted proposals, effectively counteracting the efficiency loss associated with increasing the dimension of summary statistics. In this model, Box-CD intervals were notably shorter than LR intervals while maintaining coverage.
Ricker's Model: Applied to a population dynamics model with intractable likelihoods, the method produced conformal confidence regions for both univariate and bivariate parameter settings.

4. Significance and Claims

The paper claims that Box-Confidence Depth offers a unified strategy for inference that:

Eliminates the need for scalar reduction: It allows the direct use of raw data or multivariate summary statistics without forcing them into a single scalar value, preserving information.
Operates without likelihoods: It is suitable for likelihood-free inference scenarios where the likelihood is unknown or intractable.
Provides calibrated frequentist inference: Unlike some simulation-based Bayesian approximations, it constructs confidence sets with guaranteed coverage properties under specific regularity conditions.
Handles multivariate ordering: It solves the problem of ordering in multivariate spaces by using a depth function based on hyper-rectangles, avoiding the computational complexity of simplicial or half-space depths.
Robustness: The method is shown to be invariant under order-preserving transformations of the summary statistics.

The authors position the method as a practical alternative for researchers facing complex models where traditional asymptotic theory fails and where existing likelihood-free methods struggle with high-dimensional data or require arbitrary distance thresholds.