Data and Analysis for "Multimodal Anthropomorphism in Companion Chatbots"

This repository contains the de-identified data and analysis scripts for the Master's thesis project titled
"Multimodal Anthropomorphism in Companion Chatbots: Examining Avatar Choice and Adaptive Language Style with Kagami,"
conducted at the University of Minnesota (IRB #STUDY00025677).

Preregistration

This study was preregistered on the Open Science Framework (OSF). The full, public preregistration—including the study design, hypotheses, and analysis plan—can be accessed at the following link:

• OSF Preregistration: https://osf.io/we24d

Project Summary

This study investigated how visual and linguistic anthropomorphism affect perceptions of a companion chatbot. In a 3×2 between-subjects experiment, N = 162 participants recruited from Prolific chatted with the Kagami AI for 10 minutes and completed pre- and post-task surveys. Conditions varied on avatar type (Premade, Generated, None) and linguistic adaptivity (Adaptive vs. Static Language Style Matching).

How to Reproduce All Analyses

This repository is structured for full reproducibility of all figures and tables.

Prerequisites: * You have Python 3 installed. * You have cloned this repository.

Steps:

1. Install Dependencies: Navigate to the project's root directory in your terminal and install the required Python packages.

   bash pip install -r requirements.txt

2. Run the Master Script: First, make the master script executable (you only need to do this once).

   bash chmod +x run_all.sh

   Then, execute the script to run all analyses and generate all outputs.

   bash ./run_all.sh

This will run all analysis scripts in sequence. When it's finished, all tables and figures in the /reports directory will be fully regenerated from the public data.

Repository Structure

. ├── data/ │ ├── analysis_dataset_deidentified.csv │ ├── chat_metrics_derived.csv │ ├── generated_prompts_coded_deidentified.csv │ └── content_analysis_bins_deidentified.csv ├── reports/ │ ├── figures/ │ ├── tables/ │ └── (analysis_log.txt files...) ├── src/ │ ├── confirmatory_analysis.py │ ├── content_analysis.py │ ├── exploratory_analysis.py │ ├── linguistic_audit.py │ ├── robustness_checks.py │ ├── power_sensitivity.py │ ├── data_preprocessing.py # (Documentation Only) │ ├── content_binning_methodology.py # (Documentation Only) │ ├── lsm_trajectory.py # (Documentation Only) │ └── semantic_synchrony.py # (Documentation Only) ├── materials/ │ └── (IRB Protocol, Consent Forms, etc.) ├── .gitignore ├── LICENSE └── README.md

Data Availability & Ethical Considerations

The public datasets are located in the /data/ directory and have been rigorously de-identified in accordance with the approved IRB protocol.

• analysis_dataset_deidentified.csv: The main dataset with survey responses, de-identified demographics, and experimental conditions for N=162 participants.
• chat_metrics_derived.csv: Aggregate linguistic features (e.g., LIWC scores, word counts) at the participant level.
• generated_prompts_coded_deidentified.csv: Coded themes of user-generated avatar prompts, with all raw text removed.
• content_analysis_bins_deidentified.csv: Final conversational content bin assignment for each participant.

Important Note on Withheld Data

To ensure participant confidentiality, all raw text—including full chat logs and open-ended survey responses—and granular per-turn behavioral data have been withheld from this public repository.

Analysis Scripts Overview

Reproducible Analyses

The following scripts are executed by run_all.sh and use only the public data files.

• confirmatory_analysis.py: Runs the preregistered 3x2 ANOVAs and generates Figure 3.
• content_analysis.py: Analyzes the conversational content bins and generates Figures 5, 6, and 7.
• exploratory_analysis.py: Runs exploratory moderation/mediation analyses and generates Figures 8 and 9.
• linguistic_audit.py: Conducts supplementary t-tests comparing aggregate linguistic features between conditions.
• robustness_checks.py: Performs supplementary analyses, including ANCOVA and FDR corrections.
• power_sensitivity.py: Conducts a post-hoc sensitivity power analysis.

Non-Reproducible Methodologies (Documentation Only)

These scripts contain the complete, original code from the private analysis pipeline. They document the methodology for analyses that required sensitive data but will not execute.

• data_preprocessing.py: Documents how the raw, private data was cleaned and processed into the final public datasets.
• content_binning_methodology.py: Documents the BERTopic and LIWC-based pipeline used to classify conversations into thematic bins.
• lsm_trajectory.py: Contains the code that generated Figure 4.
• semantic_synchrony.py: Contains the code that generated Figure 10.

Study Materials

The materials/ folder contains participant-facing documents, including: - Consent forms - Post-task surveys - Experimental stimuli (avatar options, LSM prompt variants)

No identifying information or internal IRB documentation is included.

License

All files in this repository are shared under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0) license.
This applies to all code, data, and study materials contained herein.

Note: The Kagami chatbot application used in the experiment is maintained in a separate repository with its own license:
🔗 https://github.com/tj-brandt/kagami

Citation

Brandt, T.J. (2025). Kagami Study – Data Analysis Repository (Version 1.0-prereg) [Data and Code]. Zenodo. https://doi.org/10.5281/zenodo.15800990