Pre- and Post-Error Dynamics: LME Analysis Pipeline

Overview

Linear mixed-effects analyses of trial-by-trial behavior around errors. Uses large-scale datasets, residual-based RT metrics, and matched error vs no-error groups to quantify pre-error speeding and post-error slowing

Directory Structure

├── 1_data_cleaning.R                    # Preprocess raw data, calculate RTs, filter users
├── 2_match_groups.R                     # Create matched error and no-error groups
├── 3_residuals_tests_plots.R            # LME residuals, per-trial tests, log-residual plots
├── 4_trial-specific_residuals_plots.R   # Residuals per trial, per-trial stats, plots
└── README.md                            # Project description and instructions

1. 1_data_cleaning.R

Purpose:
Initial preprocessing of raw Flash Challenge data. Calculates response times, performs log transformation, filters participants, and prepares trial-level variables for later analyses.

Input:

• Flash_95%_Sarah.csv: Raw 95% Flash Challenge dataset (Airport Scanner mobile game).

Cleaning Steps:

• Compute Response Times: Create Touch Time, Swipe Time, and Total RT from raw timestamps (WaitTime, DragTime, ReleaseTime).
• Log Transform RTs: Add log-transformed versions for each RT measure (Touch_Time_log, Swipe_Time_log, Total_RT_log).
• Filter Trials: Keep only Tier 1, Day 1, first challenge per user. Remove outlier trials (Total_RT < 250 ms or Total_RT > 5000 ms) and users with fewer than 8 correct trials.
• Assign Trial Numbers: Add sequential trial numbers within each user.
• Add N−1 Accuracy: Compute accuracy of the previous trial to capture sequential effects.
• Label Error Status: Mark each trial as Pre-Error, Error, Post-Error, or Other.
• Compute Total Trials: Record the total number of trials completed per user.
• Standardize RTs: Create z-scored RTs within user for Touch and Swipe times.

Output:

• FlashData_95pct.csv: Preprocessed dataset.
• only1error_95pct.csv: Subset with users who made exactly one error.
• noerror_95pct.csv: Subset with users who made no errors.

---

2. 2_match_groups.R

Purpose:
Create matched error and no-error cohorts with identical trial-count distributions, add trial-relative indices, and save clean subsets for downstream modeling.

Input:

• FlashData_95pct.csv — preprocessed dataset from 1_data_cleaning.R.

Steps:

• Identify Groups:
  • Error group: users with exactly one error.
  • No-error group: users with zero errors.
• Compute Total Trials: Count trials per user in each group.
• Match on Trial Length: Intersect available TotalTrials values and filter both cohorts to shared values only.
• Sample Users per Length: For each TotalTrials value, randomly sample equal numbers of users from both groups (reproducible seed).
• Carry Over Error Trial Index: From the error group, copy each user’s ErrorTrial to the matched no-error user with the same TotalTrials.
• Compute Trial Index: Add TrialsSinceError = TrialNumber - ErrorTrial for both groups (the no-error cohort uses the borrowed ErrorTrial index).
• Save Matched Subsets: User lists and matched trial-level data for both groups.

Output:

• subset_error_userIds.csv, subset_noerror_userIds.csv — matched user lists by TotalTrials.
• only1error_95pct_subset_users.csv — matched error cohort with TrialsSinceError.
• noerror_95pct_subset_users.csv — matched no-error cohort with TrialsSinceError.

---

3. 3_residuals_tests_plots.R

Purpose:
Fit linear mixed-effects (LME) models on log RTs, compute residuals for error vs no-error groups, run per-trial group comparisons across ±15 trials, and visualize log-residual trajectories.

Input:

• only1error_95pct_subset_users.csv — matched error cohort.
• noerror_95pct_subset_users.csv — matched no-error cohort.

Steps:

• Preprocess for Modeling:
  • Optionally remap ItemId (e.g., add 10,000 if isIllegal == 1) to separate legal/illegal items.
  • Keep finite values for Touch_Time_log, Swipe_Time_log, Total_RT_log.
• Fit LME Models (log scale):
  • Touch: Touch_Time_log ~ TrialsSinceError + (1|UserId) + (1|TrialNumber) + (1|TotalTrials) + (1|ItemId)
  • Swipe: Swipe_Time_log ~ TrialsSinceError + (1|UserId) + (1|TrialNumber) + (1|TotalTrials) + (1|ItemId)
  • Total: Total_RT_log ~ TrialsSinceError + (1|UserId) + (1|TrialNumber) + (1|TotalTrials) + (1|ItemId)
• Extract Residuals: Add model residuals per trial for each RT (Touch/Swipe/Total), for both groups.
• Combine & Window: Merge groups, add Group ∈ {Error, NoError}, keep TrialsSinceError ∈ [−15, +15].
• Per-Trial Tests: For each TrialsSinceError, run two-sample t-tests of residuals (Error vs NoError) for Touch, Swipe, and Total (using broom::tidy).
• Model-Level Coef p-values: Extract p-values for TrialsSinceError from each LME summary (Touch/Swipe/Total).
• Plots (log residuals): Line means with SE ribbons for Touch, Swipe, and Total as a function of TrialsSinceError (grouped by Error/NoError).

Output:

• pvals_touch.csv, pvals_swipe.csv, pvals_total.csv — per-trial p-values (Error vs NoError) on log residuals.
• test_stats_touch.csv, test_stats_swipe.csv, test_stats_total.csv — full per-trial test statistics (estimate, t, df, p) on log residuals.
• (Optional if saving) touch_residuals_log.png, swipe_residuals_log.png, total_residuals_log.png — log-residual plots.

---

4. 4_trial-specific_residuals_plots.R

Purpose:
Recompute residuals independently at each trial index (i.e., for each TrialsSinceError = k) using intercept-only LMEs, then compare Error vs NoError per trial and visualize residual trajectories.

Input:

• only1error_95pct_subset_users.csv — matched error group.
• noerror_95pct_subset_users.csv — matched no-error group.

Steps:

• Combine & Prepare: Bind cohorts; set Group ∈ {Error, NoError}. Optionally remap ItemId when isIllegal == 1.
• Loop Over Trial Indices: For each k ∈ [−15, +15], subset TrialsSinceError == k and fit intercept-only LMEs:
  • Touch: Touch_Time_log ~ 1 + (1|TrialNumber) + (1|ItemId)
  • Swipe: Swipe_Time_log ~ 1 + (1|TrialNumber) + (1|ItemId)
  • Total: Total_RT_log ~ 1 + (1|TrialNumber) + (1|ItemId)
    Extract residuals for each RT at trial index k.
• Bind Residual Sets: Concatenate residual frames across all k for Touch/Swipe/Total.
• Per-Trial Tests: For each k, run two-sample t-tests on residuals (Error vs NoError) for the three RTs; tidy outputs with broom.
• Plots (trial-specific residuals): Mean residual trajectories with SE ribbons across TrialsSinceError for Touch, Swipe, and Total.

Output:

• pvals_touch_trial.csv, pvals_swipe_trial.csv, pvals_total_trial.csv — per-trial p-values (Error vs NoError) using trial-specific residuals.
• test_stats_touch_trial.csv, test_stats_swipe_trial.csv, test_stats_total_trial.csv — full per-trial test statistics.

---

Contact Information

For any questions regarding this project, please contact:

Name: Sarah Malykke

Email: [email protected]