rt-attention-analysis

Processes reaction-time (RT) and accuracy data from continuous-performance / sustained-attention tasks (PVT, SART, gradCPT). Computes trailing-average RT windows, cumulative mean and standard deviation, real-time fast/slow triggering thresholds (μ±σ), lapse detection, RT variability, and bootstrap /

RT & Sustained-Attention Data Analysis

Analyze reaction-time (RT) data from continuous performance tasks (CPTs) such as the PVT, SART, or gradCPT. Implements the real-time "fast/slow triggering" procedure from deBettencourt et al. (2019) and the lapse-prediction analyses from Shelat et al. (2024).

When to use

• The user has a CSV/TSV/JSON of trial-by-trial RTs (and optionally accuracy) and wants to detect lapses, run real-time triggering, compute RT variability, or summarize sustained-attention performance.

Input format

One row per trial. Minimum column: rt (response time in ms). Optional columns: trial, acc (1/0 or correct/incorrect), freq (frequent=1, infrequent=0), block, condition.

Core algorithm (real-time triggering)

For each trial i:

1. mu, sigma = cumulative mean and standard deviation of all RTs up to trial i (initialize on the first ~80 trials).
2. trailing = mean of the last 3 RTs (i-2, i-1, i).
3. If trailing < mu - sigma → fast state (low attention) → trigger a probe.
4. If trailing > mu + sigma → slow state (high attention) → trigger a probe.
5. (Optional, experiment 2b) trigger the next trial i+1 and require the three preceding trials to be frequent and correct.

Faster trailing RTs predict worse attention; slower RTs predict better attention.

Scripts

• scripts/rt_trigger.py — pure-stdlib implementation of the triggering algorithm; writes per-trial state and fast/slow probe labels to CSV. No third-party dependencies.
• scripts/rt_analyze.py — full pipeline: load, clean, lapse detection, RT variability, Cohen's d, Spearman r, bootstrap CI, and an RT time-series plot (matplotlib optional).

Run:

bash

python scripts/rt_trigger.py data.csv --out triggered.csv
python scripts/rt_analyze.py data.csv --out report/

Key metrics

• RT variability = SD of RTs on correct frequent trials (higher ⇒ worse attention).
• Lapse = a fast trailing RT (e.g. below μ − σ) preceding an error.
• Effect size = Cohen's d; correlation = Spearman's r (non-parametric).
• Bootstrap = resample participants with replacement for CIs and p-values.

See references/methods.md for the methodological background and formulas.

Notes

• Remove implausible RTs (e.g. < 100 ms or > 3 SD) before analysis and document any exclusions.
• Statistics default to non-parametric (matches deBettencourt et al. 2019).