📊 statistics

Analyzing directional errors in spatial orientation using nonparametric circular regression with mixed covariates

This paper introduces a nonparametric circular regression framework with mixed covariates for analyzing spatial orientation errors, featuring a novel bootstrap bandwidth selection method that outperforms existing approaches in simulations and reveals condition-specific nonlinear patterns in experimental data involving blind, low-vision, and sighted participants.

Original authors: Mario Francisco-Fernández, Andrea Meilán-Vila

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mario Francisco-Fernández, Andrea Meilán-Vila

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 walking through a dark room. You know where the door is, but you can't see it. You have to rely on your footsteps, the sound of your breathing, or the feeling of the air to guess where you are and which way to turn. Sometimes you guess right; sometimes you turn a little too far left or right.

This paper is about understanding why people make those directional guesses, how sensory conditions (like being blindfolded or having noise-canceling headphones) change those guesses, and how to use math to map out those errors without making up rigid rules.

Here is the breakdown of the research in simple terms:

1. The Problem: The "Spinning Top" of Human Direction

Humans are great at knowing where they are when they can see. But when we lose our sight (or hearing), our brain has to work harder to update our position.

The Experiment: Researchers took 64 people (some sighted, some with low vision, some blind) and asked them to walk a specific path in a room. At the end, they had to point to where a target was.
The Twist: The researchers didn't just ask, "How far off were you?" They asked, "Did you turn too far clockwise or too far counter-clockwise?" This is called signed angular error. It's like knowing if you missed the target by turning left or right, not just how many degrees you missed.

2. The Challenge: The "Round" Math Problem

Standard math (like drawing a straight line on a graph) doesn't work well for directions.

The Analogy: Imagine a clock. If you are at 11:59 and you move one minute forward, you are at 12:00. But if you move one minute backward, you are at 11:59. In normal math, 11:59 and 12:01 are far apart. On a clock, they are right next to each other.
The Issue: Most statistical tools treat numbers like a straight line. They get confused by the "wrap-around" nature of a circle (0 degrees is the same as 360 degrees). The authors had to invent a new way to do statistics that respects the "roundness" of directions.

3. The Solution: A "Smart Blender" for Data

The researchers created a new statistical tool called Nonparametric Circular Regression with Mixed Covariates. That's a mouthful, so let's break it down with an analogy:

The "Smart Blender" (The Estimator): Imagine you have a blender. You throw in two types of ingredients:
1. Continuous ingredients: Things that change smoothly, like the distance to the target (3 feet, 4 feet, 5 feet).
2. Categorical ingredients: Things that are distinct categories, like the sensory condition (Blindfolded, Hearing only, Full Vision).
The Magic: This tool doesn't just average everything together. It blends the data so that it can say, "Okay, for people walking 10 feet away while blindfolded, the error looks like this curve. But for people walking 10 feet away with full vision, the error looks like that curve."
Why it's special: It doesn't force the data into a straight line or a simple curve. It lets the data tell its own story, finding complex, wiggly patterns that other methods might miss.

4. The Tricky Part: Tuning the "Focus" (Bandwidth Selection)

In this kind of math, you have to choose a "focus" level, called bandwidth.

Too much focus (Small bandwidth): The tool looks at every single data point individually. It gets very jittery and noisy, like a shaky camera.
Too little focus (Large bandwidth): The tool smooths everything out too much. It misses the important details, like a blurry photo.
The Goal: Find the "Goldilocks" focus—not too sharp, not too blurry.

The authors tested three ways to find this Goldilocks focus:

Cross-Validation: Trying to guess the answer by leaving one person out and seeing if the model predicts them correctly.
Rule-of-Thumb: A quick, simple guess based on how many people were in the study.
Bootstrap (The Winner): They created thousands of "fake" versions of the experiment by randomly resampling the real data. They tested which focus level worked best across all these fake worlds. This method turned out to be the most stable and accurate.

5. What They Found: The "Sensory Map"

When they applied their new tool to the real data, they found some fascinating patterns:

Full Vision (Control): People were very accurate. The "error curve" stayed flat and close to zero.
Blindfolded + Noise (Deprivation): This was the hardest. People didn't just get more wrong; they got wrong in a specific, systematic way. As the target got farther away, their pointing error drifted significantly in one direction.
The "Drift": The math showed that under sensory deprivation, the brain doesn't just get "noisy"; it starts to "drift" like a boat without an anchor. The error grew as the distance increased, showing a clear, non-linear pattern that simple averages would have missed.

6. Why This Matters

This isn't just about math; it's about helping people.

For Assistive Tech: If we know exactly how and when people drift when they can't see, we can build better GPS or navigation aids for the blind that correct for those specific errors.
For Rehabilitation: Therapists can understand that losing vision doesn't just make navigation "harder"; it changes the type of error people make.
For Design: Architects can design buildings with better "sensory cues" (sounds, textures) to help people navigate when lights go out or in confusing spaces.

Summary

The authors built a specialized, round-aware math tool that can mix different types of data (like distance and sensory conditions) to map out how humans get lost. They proved that using a resampling technique (Bootstrap) to tune this tool gives the best results. Their analysis revealed that when we lose our senses, our brains don't just get messy; they develop specific, predictable "drifts" that get worse the farther we go. This helps us design better tools and environments for everyone.

1. Problem Statement

The paper addresses the statistical challenge of modeling circular response data (data lying on a circle, e.g., angles) in the presence of mixed covariates (a combination of continuous and categorical predictors).

Context: The authors analyze experimental data from Legge et al. (2016b) regarding spatial orientation. Participants (sighted, low-vision, and blind) performed navigation tasks under five different sensory conditions (Control, Preview, Forward Facing, Auditory, Deprivation).
The Response: The outcome is the signed angular error (the difference between the reported and true target direction), which is circular data. Standard linear regression is inappropriate because it ignores the periodic nature of angles (e.g., $0^\circ$ and $360^\circ$ are identical).
The Challenge: Existing nonparametric circular regression methods typically handle only continuous covariates. There is a lack of practical methodology for:
1. Handling mixed covariates (continuous task demand + categorical sensory conditions) in a circular setting.
2. Selecting optimal bandwidths (smoothing parameters) for these mixed models.
3. Quantifying uncertainty via simultaneous confidence bands for circular functions.

2. Methodology

The authors propose a nonparametric circular regression framework based on a product-kernel Nadaraya-Watson (NW) estimator.

A. The Model and Estimator

Model: $\Theta_i = [m(X_i, Z_i) + \varepsilon_i] \pmod{2\pi}$ , where $\Theta$ is the circular response, $X$ are continuous covariates, and $Z$ are categorical covariates.
Objective: Estimate the regression function $m(x, z)$ , defined as the minimizer of the angular risk $E\{1 - \cos[\Theta - m(X, Z)]\}$ .
Estimator: The authors use a product-kernel NW estimator:
$\hat{m}_H(x, z) = \text{atan2}(\hat{m}_{1,H}(x, z), \hat{m}_{2,H}(x, z))$
where $\hat{m}_{1,H}$ and $\hat{m}_{2,H}$ are weighted averages of $\sin(\Theta)$ and $\cos(\Theta)$ , respectively.
Weights: The weights combine a continuous kernel (e.g., Gaussian) for $X$ and a categorical kernel (e.g., Aitchison–Aitken) for $Z$ . This allows for joint smoothing without assuming independence between covariates.

B. Theoretical Properties

The authors derive the asymptotic bias and variance of the estimator under regularity conditions:

Bias: Decomposes into an $O(h^2)$ term (from continuous smoothing) and an $O(\lambda)$ term (from categorical smoothing). The categorical bias depends on the Hamming distance between category levels.
Variance: Of order $O(1/nh^k)$ , scaled by the local density and the residual dispersion.
Significance: These results extend existing circular regression theory to mixed covariates, providing the theoretical foundation for bandwidth selection and inference.

C. Bandwidth Selection Strategies

Since closed-form plug-in selectors are infeasible due to complex unknown quantities, the authors propose and compare three data-driven methods:

Leave-One-Out Cross-Validation (CV): Minimizes the circular prediction loss.
Rule-of-Thumb (RoT): A computationally efficient heuristic based on sample size and marginal scales (analogous to normal-reference rules).
Bootstrap Selection (Proposed): A novel procedure tailored to the circular cosine loss. It uses a residual bootstrap to estimate the global prediction risk, offering improved stability over CV.

D. Simultaneous Confidence Bands

To quantify uncertainty globally, the authors develop a bootstrap algorithm to construct simultaneous confidence bands:

It generates pseudo-samples via residual resampling.
It uses an iterative calibration procedure to determine the critical quantile that ensures the band covers the true function simultaneously across a grid of predictor values with a specified confidence level ( $1-\alpha$ ).

3. Key Contributions

Methodological Framework: Development of a unified nonparametric circular regression estimator for mixed continuous and categorical covariates.
Theoretical Advancement: Derivation of asymptotic bias and variance expressions for the mixed-covariate circular NW estimator.
Practical Tools:
- Proposal of a bootstrap bandwidth selector specifically designed for circular data, which outperforms CV and RoT in simulations.
- Development of simultaneous confidence bands for circular regression functions.
Software Implementation: All methods are implemented in the R package circMixedReg.

4. Results

A. Simulation Study

The authors conducted a simulation study comparing the three bandwidth selectors (CV, Bootstrap, RoT) across various sample sizes ( $n=100, 200, 500$ ) and noise levels.

Performance Metric: Circular Average Squared Error (CASE) against a known Monte Carlo benchmark.
Findings:
- The Bootstrap selector consistently achieved the lowest average error and the most stable bandwidth choices (closest to the optimal benchmark).
- CV showed higher variability and sensitivity to sample noise.
- RoT was computationally fast but tended to oversmooth (higher bias), though it served as a robust baseline.
- The Bootstrap method offered the best bias-variance trade-off, particularly for complex regression functions and smaller sample sizes.

B. Real-Data Application (Spatial Orientation)

The framework was applied to the Legge et al. (2016b) dataset to analyze how directional error varies with target distance (continuous) and sensory condition (categorical).

Exploratory Analysis: Raw data showed that sensory deprivation leads to increased dispersion and systematic directional bias.
Regression Findings:
- Control Condition: Errors remained close to zero across all distances with tight confidence bands.
- Reduced Sensory Conditions (Auditory, Deprivation, etc.): The regression curves revealed nonlinear, distance-dependent patterns. For example, in the "Auditory" condition, errors drifted positively as distance increased. In "Deprivation," there was a distinct positive shift and wider uncertainty.
- Simultaneous Bands: The bands confirmed that for certain distance ranges, the mean error significantly deviated from zero (no-error line), providing statistical evidence of systematic bias under specific sensory constraints.
Extension: A two-predictor model (target distance + distance-estimation error) showed that including perceptual bias improved model fit, capturing more complex interactions between physical demand and subjective perception.

5. Significance and Implications

Cognitive Science & Rehabilitation: The study provides a rigorous statistical tool to quantify how sensory restrictions (e.g., blindness, noise) affect spatial updating. It reveals that errors are not just random noise but structured, nonlinear biases dependent on task demand. This informs the design of assistive technologies and rehabilitation strategies.
Statistical Innovation: The paper fills a critical gap in circular data analysis by providing a practical, theoretically grounded method for mixed covariates. The bootstrap bandwidth selector and simultaneous bands offer robust tools for inference in complex circular datasets.
Generalizability: The framework is applicable beyond spatial orientation to any domain involving circular responses and mixed predictors (e.g., wind direction analysis with categorical weather types, or time-of-day analysis with categorical event types).

In conclusion, the paper successfully bridges the gap between theoretical circular statistics and practical application, offering a flexible, data-driven framework for analyzing directional errors under diverse sensory conditions.