stratdyn

A web application used to run a human-subjects behavioral experiment on collaborative vs. individual decision-making under risk. Participants are paired up and repeatedly choose between a set of "designs" for a shared task, each with different upside/downside payoffs, and can pursue a collaborative or an individual strategy. The app also supports an optional AI decision aid ("the robot") and mediator information panel that can be shown to participants during the task.

This repository contains the experiment server/client code and the raw data collected across control and treatment study sessions. It accompanies a manuscript reporting the study's results.

Study flow

Each session is driven by an admin (who advances/returns the shared task index for all connected participants) and proceeds through the following screens, in order, for each participant:

1. Wait screen — before the session admin starts the study.
2. Demographics survey
3. Pre-task survey
4. Design tasks — a sequence of partnered decision rounds. In each round a participant:
   • sees a task with several design options, each with an upside and a downside payoff (see data/experiment.json),
   • optionally consults the robot recommendation and/or mediator info,
   • picks a design and states whether their strategy was collaborative or individual,
   • reports their belief about the probability their partner will act collaboratively (collabBelief),
   • receives a score based on their choice and their partner's choice for that round (see the payoff logic in stratdyn.js, submit-decision).
5. Post-task survey
6. Thank you screen

Decisions and survey responses are appended live to CSV files in results/ as participants progress (see results/README.md for the data dictionary).

Repository layout

app.js                  Express app setup (static file serving, middleware)
bin/www                 Server entry point (creates the HTTP server)
stratdyn.js             Core experiment logic: socket.io event handlers,
                         session state machine, and CSV logging
data/
  experiment.json       Task definitions, upside/downside payoffs, and the
                         partner/task assignment for each participant
  adminCredentials.json Login credentials for the session admin
  userCredentials.json  Login credentials for study participants
public/                 Client-side HTML/CSS/JS for the experiment UI
  index.html            The experiment UI (welcome/survey/task/admin screens)
  index.js              Client-side socket.io event handling and UI logic
results/                Collected data, one subfolder per study arm
  control/              Control-group sessions
  treatment/             Treatment-group ("experimental group") sessions
analysis/               Scripts for turning results/ into analysis-ready tables
  build_survey_datatable.py  Merges demographics/presurvey/postsurvey into
                              one per-participant survey_data.csv
  build_task_datatable.py    Merges each task round's two partner rows into
                              one per-round task_data.csv
  build_task_summary.py      Derives one payoff-structure summary row per
                              task index from data/experiment.json

Running locally

npm install
npm start

This starts the server on the port configured in bin/www (default 3000). Participants and the admin log in from public/index.html using the credentials in data/adminCredentials.json / data/userCredentials.json. The admin account sees a live dashboard of connected participants and their in-progress decisions, and uses it to advance the group through the task sequence.

Each run of the server writes to a fixed set of output filenames, following the same {type}_{arm}_{N}.csv convention as results/ (currently a placeholder, arm/00, near the top of stratdyn.js). Update those filenames for the session about to be run, and move the previous session's output files into results/ before starting a new one, so they are not overwritten.

Data

Collected survey and task data live under results/, organized by study arm (control/, treatment/) and then by session number. See results/README.md for a description of each file type and its columns.

analysis/build_survey_datatable.py reads every demographics_*.csv, presurvey_*.csv, and postsurvey_*.csv under results/ and merges them into one row per participant (analysis/survey_data.csv), adding arm and session columns. Run it with python analysis/build_survey_datatable.py after any change to results/.

analysis/build_task_datatable.py reads every task_*.csv and merges each round's two per-partner rows into one row per round (analysis/task_data.csv), with each partner's task, design, strategy, collaboration belief, robot usage, and score in separate _1/_2 columns. The four Training Task rounds every pair starts with, plus six "distraction" tasks that lacked the study's target payoff dynamic (Task Idoha, Task Florida, Task Utah, Task Massachusetts, Task Montana, Task Mississippi), are excluded, leaving 30 rounds per pair; task_1/task_2 hold the task's index into data/experiment.json rather than its label. Run it with python analysis/build_task_datatable.py after any change to results/. See results/README.md for how the two raw rows per round are paired and scored.

analysis/build_task_summary.py reads data/experiment.json's tasks array and derives one payoff-structure summary row per task index (analysis/task_summary.csv), including each task's paired task index and its 6×5 difficulty/payoff-magnitude factorial grouping. See analysis/README.md for the full column reference.

Known issues

• The demographics survey's client-side submit handler had a copy-paste bug (public/index.js) that caused the q4 and q5 answers to be recorded as duplicates of the q3 answer in every session collected so far. This has been fixed in the code, but the historical CSVs in results/ still reflect the bug — see results/README.md for details before using those two columns.
• A subset of participants were assigned task index 23 twice and never saw task index 33, due to a data-entry error (33 was mistakenly entered as 23) when the assignment sequences were generated. This is reflected in both the raw task_*.csv files and the reconstructed assignments in data/experiment.json — see data/README.md for details.