@brainpilot/skills 0.0.6

package/skills/04_Psycholinguistics/self-paced-reading-designer/SKILL.md

@@ -0,0 +1,257 @@
+---
+name: "self-paced-reading-designer"
+description: "Expert guidance for designing self-paced reading experiments: region segmentation, timing parameters, comprehension probes, and spillover analysis"
+domain: "psycholinguistics"
+version: "1.0.0"
+authors:
+ - "AI-generated with domain expert review"
+papers:
+ - "Just, Carpenter, & Woolley, 1982"
+ - "Jegerski, 2014"
+ - "Keating & Jegerski, 2015"
+ - "Mitchell, 2004"
+ - "Boyce, Futrell, & Levy, 2020"
+dependencies:
+ required:
+ - research-literacy
+review_status: "ai-generated"
+---
+
+# Self-Paced Reading Designer
+
+This skill encodes expert knowledge for designing self-paced reading (SPR) experiments in psycholinguistics. SPR is the most widely used behavioral method for studying real-time sentence comprehension during reading (Jegerski, 2014). A competent programmer without psycholinguistics training will reliably make errors in region segmentation, spillover design, and comprehension question construction -- all of which invalidate the resulting data.
+
+For detailed region segmentation strategies, see `references/region-segmentation.md`.
+For statistical analysis guidance, see `references/analysis-guide.md`.
+
+---
+
+## Why SPR Design Requires Domain Expertise
+
+Self-paced reading appears deceptively simple: participants press a button to reveal successive words. But the scientific value of an SPR experiment depends entirely on decisions that require psycholinguistic training:
+
+1. **Region boundaries determine what you can measure.** A critical region that spans a clause boundary conflates syntactic processing with wrap-up effects (Just & Carpenter, 1980). A non-specialist would not know this.
+2. **Spillover is not a bug -- it is the primary data pattern.** In SPR, processing difficulty at word N often appears in reading times at words N+1 and N+2, not at word N itself (Mitchell, 2004; Rayner, 1998). Failing to include and analyze spillover regions means missing the effect entirely.
+3. **Comprehension questions that target the critical manipulation create demand characteristics.** Participants learn to attend strategically to the manipulation, distorting natural reading patterns (Jegerski, 2014).
+4. **Word length and frequency confounds are invisible to non-specialists.** If the critical word in condition A is longer or less frequent than in condition B, reading time differences reflect lexical properties, not the intended manipulation (Keating & Jegerski, 2015).
+
+---
+
+## Research Planning Protocol
+
+Before executing the domain-specific steps below, you MUST:
+
+1. **State the research question** — What specific sentence processing question is this SPR study addressing?
+2. **Justify the method choice** — Why SPR (not eye-tracking, ERP, acceptability judgment)? What alternatives were considered?
+3. **Declare expected outcomes** — What reading time pattern (at which region) would support vs. refute the hypothesis?
+4. **Note assumptions and limitations** — What does SPR assume? Where could it mislead (e.g., lack of regressive eye movements)?
+5. **Present the plan to the user and WAIT for confirmation** before proceeding.
+
+For detailed methodology guidance, see the `research-literacy` skill.
+
+
+## ⚠️ Verification Notice
+
+This skill was generated by AI from academic literature. All parameters, thresholds, and citations require independent verification before use in research. If you find errors, please [open an issue](https://github.com/HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills/issues).
+
+## Core Workflow
+
+### Step 1: Select a Presentation Method
+
+Choose based on your research question, population, and resources:
+
+#### 1A. Non-Cumulative Moving Window (Standard)
+
+- The sentence is displayed as dashes; each button press reveals the next word and re-masks the previous one (Just, Carpenter, & Woolley, 1982)
+- **Advantages**: Most widely used, large existing literature for comparison, preserves spatial layout information
+- **Disadvantages**: Prevents regressions (unlike natural reading), produces spillover effects that spread over 2-4 words (Mitchell, 2004)
+- **Use when**: You need comparability with the existing SPR literature; you are studying incremental sentence processing
+
+#### 1B. Cumulative Moving Window
+
+- Each button press reveals the next word, but previously revealed words remain visible
+- **Advantages**: More similar to natural reading (partial text context remains)
+- **Disadvantages**: Rarely used; harder to compare with the dominant non-cumulative literature; participants may re-read prior context, introducing noise (Jegerski, 2014)
+- **Use when**: Naturalness of reading is more important than comparability with prior work
+
+#### 1C. Phrase-by-Phrase Presentation
+
+- Sentences are segmented into multi-word regions; each button press reveals a phrase
+- **Advantages**: Faster for participants; appropriate when word-level resolution is not needed
+- **Disadvantages**: Region boundaries must be linguistically principled (see `references/region-segmentation.md`); reduces temporal resolution; risks confounding region length with reading time
+- **Use when**: Your manipulation spans a multi-word constituent and word-by-word resolution is unnecessary
+
+#### 1D. Centered (RSVP-style) Presentation
+
+- Words appear one at a time at a fixed screen position (typically center)
+- **Advantages**: Eliminates eye movement confounds; simpler programming
+- **Disadvantages**: Destroys spatial layout; removes positional information that readers normally use; rarely used in modern SPR (Jegerski, 2014)
+- **Avoid unless**: You have a specific theoretical reason to eliminate spatial layout
+
+#### 1E. Maze Task (Modern Alternative)
+
+- Two words appear simultaneously; the participant selects the word that continues the sentence (Forster, Guerrera, & Elliot, 2009; Witzel, Witzel, & Forster, 2012)
+- **L-maze (Lexicality maze)**: Distractor is a pronounceable nonword
+- **G-maze (Grammaticality maze)**: Distractor is a real word that is ungrammatical in context
+- **A-maze (Auto-maze)**: Distractors generated automatically via NLP (Boyce, Futrell, & Levy, 2020)
+- **Advantages**: Dramatically reduced spillover compared to SPR; forced incremental processing; works well for web-based data collection (Boyce et al., 2020); better statistical power per item than SPR for syntactic effects (Witzel et al., 2012)
+- **Disadvantages**: Slower overall pace; dual-task demand (comprehension + selection); less natural than button-press SPR; requires distractor generation
+- **Use when**: You need precise localization of effects, want to reduce spillover, or plan web-based data collection
+
+### Step 2: Configure Timing Parameters
+
+| Parameter | Recommended Value | Rationale |
+|---|---|---|
+| **Response timeout** | None (self-paced) or **3000-5000 ms** per region | No timeout is standard for in-lab SPR; timeout prevents excessively slow responses in web-based studies (Boyce et al., 2020) |
+| **Inter-stimulus interval (ISI)** | **0 ms** for non-cumulative moving window | Standard practice; the next word appears immediately when the previous is masked (Just et al., 1982) |
+| **ISI for phrase-by-phrase** | **0 ms** (typical) | Any nonzero ISI introduces a blank that disrupts reading and may introduce strategic pausing |
+| **Pre-sentence fixation** | **+ or * for 500-1000 ms** | Orients attention to display location; standard in SPR (Jegerski, 2014) |
+| **Post-sentence delay** | **0-500 ms** before comprehension question | Brief delay prevents motor interference between last word button-press and question response |
+| **Practice trials** | **6-10 items minimum** | Familiarizes participants with button-press rhythm and comprehension questions; use different sentences than experimental items (Jegerski, 2014; Keating & Jegerski, 2015) |
+
+### Step 3: Design Critical Regions
+
+This is the most consequential design decision in an SPR experiment. See `references/region-segmentation.md` for full guidelines.
+
+#### Core Principles
+
+1. **Match critical regions across conditions for word length (in characters) and lexical frequency.** If your manipulation requires different words, match them on length (+/- 1 character) and log frequency (use SUBTLEX-US; Brysbaert & New, 2009). Unmatched items introduce confounds that mimic or mask experimental effects.
+
+2. **Include at least 2-3 spillover words after the critical region.** Processing difficulty at the critical region reliably spills over to subsequent words in SPR (Just et al., 1982; Mitchell, 2004; Rayner, 1998). Without spillover regions, you will miss your effect. These spillover words must be identical across conditions.
+
+3. **Avoid placing critical regions at clause or sentence boundaries.** Reading times at clause-final and sentence-final positions are inflated by wrap-up processes -- integration of clause-level meaning, discourse updating, and possibly implicit prosodic boundary effects (Just & Carpenter, 1980; Warren, White, & Reichle, 2009). This inflation is independent of your manipulation and adds noise.
+
+4. **Keep critical regions short (ideally a single word).** Multi-word critical regions reduce temporal resolution and introduce length confounds. If you must use a multi-word region, it must have the same number of words and matched total character length across conditions.
+
+5. **Ensure the pre-critical region is identical across conditions.** Any difference before the critical word can create baseline differences in reading time that propagate into the critical region via spillover.
+
+### Step 4: Design Comprehension Questions
+
+Comprehension questions serve two purposes: ensuring participants read for meaning, and providing an exclusion criterion for inattentive participants.
+
+#### Guidelines
+
+| Parameter | Recommendation | Rationale |
+|---|---|---|
+| **Proportion of trials with questions** | **1/3 to 1/2** of all trials (experimental + filler) | Fewer than 1/3: participants may stop reading carefully; more than 1/2: task becomes tedious, and participants may shift to a question-anticipation strategy (Just et al., 1982; Jegerski, 2014) |
+| **Answer balance** | **50% yes / 50% no** for yes/no questions | Prevents response bias toward one answer |
+| **Question content** | Target semantic content of the sentence, NOT the critical manipulation | Questions about the manipulation teach participants what you are studying, inducing strategic reading (Jegerski, 2014) |
+| **Accuracy exclusion threshold** | **>80%** correct to retain participant | Standard criterion; lower accuracy suggests the participant was not reading for comprehension (Jegerski, 2014; common practice across SPR studies) |
+| **Question timing** | Immediately after the sentence (or after the final button press) | Delayed questions test memory, not comprehension |
+
+#### Example of Good vs. Bad Comprehension Questions
+
+Suppose the experimental sentence manipulates relative clause attachment:
+
+> *The maid of the actress who was on the balcony shouted to the crowd.*
+
+- **Good question**: "Did someone shout to the crowd?" (targets overall meaning, not the critical attachment)
+- **Bad question**: "Who was on the balcony?" (directly probes the ambiguity under investigation, alerting participants to the manipulation)
+
+### Step 5: Design Item and Condition Structure
+
+#### Latin Square Design
+
+For within-subjects manipulations, use a Latin square design so that each participant sees each item in exactly one condition, and each condition is seen equally often across participants (Keating & Jegerski, 2015).
+
+- For a 2-condition design: 2 lists; each item appears in condition A for half the participants, condition B for the other half
+- For a 2x2 design: 4 lists (one per condition combination)
+- Assign participants to lists in rotation
+
+#### Items Per Condition
+
+| Population | Minimum Items per Condition | Rationale |
+|---|---|---|
+| **L1 speakers, robust effect** (e.g., garden-path) | **24** items per condition | Sufficient for medium-to-large effects in mixed models (Keating & Jegerski, 2015) |
+| **L1 speakers, subtle effect** (e.g., pragmatic inference) | **32-40** items per condition | Smaller effects require more items for adequate power (Keating & Jegerski, 2015; Brysbaert & Stevens, 2018) |
+| **L2 speakers** | **32-40** items per condition | Higher variability in L2 populations requires more observations (Marsden, Thompson, & Plonsky, 2018) |
+
+#### Filler Items
+
+| Parameter | Recommendation | Rationale |
+|---|---|---|
+| **Filler-to-experimental ratio** | **2:1 or 3:1** (fillers : experimental items) | Prevents participants from identifying the experimental pattern; higher ratios reduce strategic processing (Keating & Jegerski, 2015) |
+| **Filler variety** | Include multiple sentence types, lengths, and structures | Monotonous fillers fail to mask the experimental manipulation |
+| **Filler complexity** | Include some fillers of similar complexity to experimental items | If only experimental items are complex, participants learn to attend differently to them |
+| **Comprehension questions on fillers** | Yes -- at least the same rate as on experimental items | If questions only follow experimental items, participants learn that complex sentences predict questions |
+
+### Step 6: Decide Between SPR and Eye-Tracking
+
+This is a design-level decision that should be made before programming the experiment.
+
+| Criterion | SPR | Eye-Tracking |
+|---|---|---|
+| **Equipment cost** | Low (any computer) | High (dedicated eye-tracker, ~$20,000-$50,000) |
+| **Online data collection** | Yes (web-based SPR and Maze work well) | No (requires in-lab calibration) |
+| **Temporal resolution** | Word-by-word, with substantial spillover | Multiple fixation measures (first fixation, gaze duration, go-past, total time, regressions) |
+| **Regressions** | Not measurable (non-cumulative display prevents rereading) | Yes -- regressions are a primary measure of reanalysis |
+| **Ecological validity** | Moderate (button-press is unnatural, but spatial layout preserved) | Higher (closer to natural reading) |
+| **Sensitivity to early/late processing stages** | Low (only a single RT per region, which blends all processing stages) | High (first-pass vs. second-pass measures separate early from late processing; Rayner, 1998) |
+| **Best for** | Robust syntactic/semantic effects, web-based or underfunded studies, L2 populations without lab access | Nuanced temporal dynamics, distinguishing processing stages, studying regressions, garden-path recovery |
+
+**Rule of thumb**: If you only need to know *whether* a manipulation affects reading time, SPR is sufficient. If you need to know *when* during processing the effect occurs (early lexical access vs. late reanalysis), use eye-tracking.
+
+---
+
+## Common Pitfalls
+
+These are errors that non-specialists routinely make:
+
+1. **No spillover region.** The most common fatal flaw. If the sentence ends at or immediately after the critical word, spillover effects have nowhere to appear, and the effect is lost. Always include 2-3 words of identical post-critical material across conditions.
+
+2. **Critical region at a clause boundary.** Wrap-up effects at clause-final positions (Just & Carpenter, 1980) inflate reading times by **50-100+ ms** regardless of condition, swamping the experimental effect or producing spurious interactions.
+
+3. **Length/frequency mismatch.** Longer words take approximately **30-40 ms per additional character** in SPR (Ferreira & Clifton, 1986). A 2-character difference between conditions creates a ~60-80 ms confound, which can easily exceed the size of most psycholinguistic effects.
+
+4. **Comprehension questions targeting the manipulation.** This transforms the experiment from measuring natural reading into measuring strategic disambiguation. Participants adapt within 10-15 trials (Jegerski, 2014).
+
+5. **Too few items per condition.** With fewer than 24 items per condition, even large effects (d = 0.8) may not reach significance in mixed-effects models, particularly with by-item random slopes (Brysbaert & Stevens, 2018).
+
+6. **No fillers or insufficient fillers.** Without a 2:1 filler-to-item ratio, participants identify the experimental manipulation and shift to strategic reading (Keating & Jegerski, 2015).
+
+7. **Analyzing only the critical region.** Even when an effect appears on the critical word, it typically continues into the spillover region. Analyzing only one region provides an incomplete picture and may miss effects that appear exclusively in spillover.
+
+8. **Using raw reading times without controlling for word length.** Raw RTs conflate lexical processing speed with the experimental manipulation. Either match word length precisely or use residual RTs / include word length as a covariate in the statistical model (Ferreira & Clifton, 1986).
+
+9. **Ignoring trial position effects.** Reading speed increases across the experiment as participants become practiced. Include trial order as a covariate or present items in a randomized order (Jegerski, 2014).
+
+10. **Not checking comprehension accuracy before analyzing RTs.** Participants with low accuracy (<80%) may not be reading for comprehension. Their RT data are uninterpretable and should be excluded (Jegerski, 2014).
+
+---
+
+## Quick Reference: SPR Design Checklist
+
+Before running your experiment, verify:
+
+- [ ] Critical regions are matched for word length and frequency across conditions
+- [ ] At least 2-3 identical spillover words follow the critical region in all conditions
+- [ ] Critical region is not at a clause or sentence boundary
+- [ ] Pre-critical region is identical across conditions
+- [ ] At least 24 items per condition (32+ for subtle effects or L2 populations)
+- [ ] Filler-to-experimental ratio is at least 2:1
+- [ ] Comprehension questions on 1/3 to 1/2 of trials, balanced yes/no
+- [ ] No comprehension question directly targets the experimental manipulation
+- [ ] Latin square counterbalancing across lists
+- [ ] 6-10 practice trials with different sentences than experimental items
+- [ ] Analysis plan includes spillover regions (at least critical +1, +2)
+- [ ] RT trimming criteria defined a priori (see `references/analysis-guide.md`)
+
+---
+
+## References
+
+- Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. *Journal of Memory and Language, 59*, 390-412.
+- Boyce, V., Futrell, R., & Levy, R. P. (2020). Maze Made Easy: Better and easier measurement of incremental processing difficulty. *Journal of Memory and Language, 111*, 104082.
+- Brysbaert, M., & New, B. (2009). Moving beyond Kucera and Francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for American English. *Behavior Research Methods, 41*, 977-990.
+- Brysbaert, M., & Stevens, M. (2018). Power analysis and effect size in mixed effects models. *Journal of Cognition, 1*(1), 9.
+- Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. *Journal of Memory and Language, 25*, 348-368.
+- Forster, K. I., Guerrera, C., & Elliot, L. (2009). The maze task: Measuring forced incremental sentence processing time. *Behavior Research Methods, 41*, 163-171.
+- Jegerski, J. (2014). Self-paced reading. In J. Jegerski & B. VanPatten (Eds.), *Research methods in second language psycholinguistics*. Routledge.
+- Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to comprehension. *Psychological Review, 87*, 329-354.
+- Just, M. A., Carpenter, P. A., & Woolley, J. D. (1982). Paradigms and processes in reading comprehension. *Journal of Experimental Psychology: General, 111*, 228-238.
+- Keating, G. D., & Jegerski, J. (2015). Experimental designs in sentence processing research. *Studies in Second Language Acquisition, 37*, 1-32.
+- Marsden, E., Thompson, S., & Plonsky, L. (2018). A methodological synthesis of self-paced reading in second language research. *Applied Psycholinguistics, 39*, 861-904.
+- Mitchell, D. C. (2004). On-line methods in language processing: Introduction and historical review. In M. Carreiras & C. Clifton (Eds.), *The on-line study of sentence comprehension*. Psychology Press.
+- Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. *Psychological Bulletin, 124*, 372-422.
+- Warren, T., White, S. J., & Reichle, E. D. (2009). Investigating the causes of wrap-up effects: Evidence from eye movements and E-Z Reader. *Cognition, 111*, 132-137.
+- Witzel, N., Witzel, J., & Forster, K. (2012). Comparisons of online reading paradigms: Eye tracking, moving-window, and maze. *Journal of Psycholinguistic Research, 41*, 105-128.

package/skills/04_Psycholinguistics/self-paced-reading-designer/references/analysis-guide.md

@@ -0,0 +1,356 @@
+# Statistical Analysis Guide for Self-Paced Reading Data
+
+This reference covers the statistical analysis pipeline specific to self-paced reading (SPR) data, from raw reading times to final inference. SPR data have distinctive properties (right-skewed distributions, nested structure, spillover autocorrelation) that require analysis approaches different from standard cognitive tasks.
+
+---
+
+## 1. Data Cleaning Pipeline
+
+Apply these steps in order before any statistical analysis.
+
+### 1.1 Exclude Participants Based on Comprehension Accuracy
+
+**Criterion**: Exclude participants with overall comprehension question accuracy **below 80%** (Jegerski, 2014; common practice across the SPR literature).
+
+**Rationale**: Low accuracy indicates the participant was not reading for comprehension. Their reading times are not interpretable as measures of linguistic processing.
+
+**Implementation**:
+- Calculate accuracy across all comprehension questions (experimental + filler)
+- Log the number of excluded participants and their accuracy rates
+- Report the exclusion criterion and how many participants were removed
+
+### 1.2 Exclude Trials with Incorrect Comprehension Responses
+
+**Convention**: For experimental items that have a comprehension question, exclude the reading time data from trials where the participant answered incorrectly.
+
+**Rationale**: An incorrect answer suggests the participant may not have been attending to the sentence on that trial. Including those RTs adds noise (Jegerski, 2014).
+
+**Note**: This step is debated. Some researchers argue that excluding incorrect trials biases the sample toward easier items. If you choose not to exclude incorrect trials, report this decision and consider analyzing accuracy as a separate dependent variable.
+
+### 1.3 Apply Absolute RT Cutoffs
+
+Remove reading times that are implausibly fast or slow. These reflect motor errors (accidental button presses) or distraction (forgetting to press), not linguistic processing.
+
+| Cutoff | Value | Rationale |
+|---|---|---|
+| **Lower bound** | **100 ms** | Below 100 ms, it is physically impossible to have read and processed a word; this reflects an accidental button press (Rayner, 1998; Jegerski, 2014) |
+| **Upper bound** | **2000-3000 ms** | Above 2000-3000 ms for a single word, the participant was likely distracted or disengaged. The exact value depends on population: use 2000 ms for L1 adults, 2500-3000 ms for L2 or clinical populations (Jegerski, 2014; Marsden, Thompson, & Plonsky, 2018) |
+
+**Important**: Define these cutoffs a priori (before looking at the data). Report the cutoffs and the percentage of data removed. Removing more than **5-10%** of data points suggests a problem with the task or participant population (Ratcliff, 1993).
+
+### 1.4 Apply SD-Based Trimming (Optional, Used with ANOVAs)
+
+If using traditional ANOVA rather than mixed-effects models, apply SD-based trimming after absolute cutoffs:
+
+| Method | Procedure | Typical Threshold |
+|---|---|---|
+| **By-participant, by-condition** | For each participant in each condition, remove RTs beyond +/- **2.5 SD** from that participant's condition mean | 2.5 SD (Ratcliff, 1993) |
+| **By-participant only** | For each participant, remove RTs beyond +/- **2.5 SD** from their overall mean | 2.5 SD |
+
+**Note on mixed-effects models**: When using linear mixed-effects models (LMMs), aggressive SD-based trimming is often unnecessary because LMMs are more robust to outliers than ANOVAs. If using LMMs, apply absolute cutoffs (Step 1.3) and consider log-transforming RTs (Step 2) instead of SD trimming. Alternatively, remove data points with residuals exceeding +/- 2.5 SD from the fitted model (Baayen, Davidson, & Bates, 2008).
+
+---
+
+## 2. RT Transformation
+
+Raw reading times are right-skewed (bounded at zero, with a long positive tail). This violates the normality assumption of linear models.
+
+### 2.1 Log-Transformation
+
+**Recommended as the default**: Apply natural log (ln) or log base 10 transformation to all reading times.
+
+```
+log_RT = log(RT)
+```
+
+**Advantages**:
+- Reduces right skew, making residuals more normally distributed
+- Converts multiplicative effects to additive ones (appropriate because many reading time effects are proportional, not absolute)
+- Standard in the SPR literature (Baayen et al., 2008)
+
+**Disadvantage**: Effects are on the log scale, making raw millisecond interpretation less intuitive. Report back-transformed means for descriptive statistics.
+
+### 2.2 Inverse (Reciprocal) Transformation
+
+**Alternative**: Use -1000/RT (negative reciprocal scaled by 1000).
+
+```
+inv_RT = -1000 / RT
+```
+
+**Advantages**:
+- Often produces better normality than log transformation for reading time data (Box & Cox, 1964)
+- Recommended by some psycholinguists (e.g., Vasishth & Nicenboim, 2016)
+
+**Disadvantage**: Even less intuitive than log RT. Negative values mean that larger (less negative) values correspond to faster reading.
+
+### 2.3 No Transformation (With Caveats)
+
+If using generalized linear mixed models (GLMMs) with a log link or Gamma family, no transformation of the raw RTs is needed because the model handles the skew internally.
+
+```
+glmer(RT ~ condition + (1 | subject) + (1 | item), family = Gamma(link = "identity"), data = d)
+```
+
+This approach is increasingly recommended (Lo & Andrews, 2015) but requires more computational resources and expertise to implement correctly.
+
+---
+
+## 3. Statistical Models
+
+### 3.1 Linear Mixed-Effects Models (Recommended)
+
+Use linear mixed-effects models with crossed random effects for subjects and items (Baayen, Davidson, & Bates, 2008). This is the current standard for SPR analysis.
+
+#### Minimal Model Structure
+
+```r
+library(lme4)
+
+model <- lmer(log_RT ~ condition +
+ (1 + condition | subject) +
+ (1 + condition | item),
+ data = data)
+```
+
+#### Random Effects Structure
+
+The random effects structure is critical and often a source of error:
+
+| Structure | When to Use | Reference |
+|---|---|---|
+| **Maximal** (random intercepts + slopes for all fixed effects, by subject and item) | Default recommendation; prevents anti-conservative p-values | Barr, Levy, Scheepers, & Tily (2013) |
+| **Parsimonious** (remove random correlations first, then slopes with near-zero variance) | When maximal model fails to converge | Bates, Kliegl, Vasishth, & Baayen (2015); Matuschek, Kliegl, Vasishth, Baayen, & Bates (2017) |
+| **Intercepts only** | Only as a last resort; may be anti-conservative | Baayen et al. (2008) |
+
+**Common convergence issues**: SPR data often produce convergence warnings with maximal models. Follow this simplification hierarchy:
+1. Remove random correlations (use `||` instead of `|` in lme4 formula)
+2. Remove random slopes with the smallest variance estimate
+3. If still non-convergent, use a different optimizer (e.g., `bobyqa` with increased iterations)
+
+#### Covariates to Include
+
+| Covariate | Why | How to Include |
+|---|---|---|
+| **Word length (characters)** | Longer words take longer to read (~30-40 ms/character; Ferreira & Clifton, 1986) | `scale(nchar)` as fixed effect |
+| **Log word frequency** | Less frequent words are slower (~20-40 ms per log unit; Rayner, 1998) | `scale(log_freq)` as fixed effect |
+| **Trial order** | Reading speed increases with practice across the experiment | `scale(trial_number)` as fixed effect |
+| **Previous word RT** | Controls for spillover autocorrelation | `scale(lag_RT)` as fixed effect (but see Section 5) |
+| **Word position in sentence** | First and last words have inflated RTs | Include as fixed effect or exclude those positions |
+
+### 3.2 Traditional ANOVA (By-Subjects and By-Items)
+
+If using ANOVA, you must run two separate analyses:
+
+1. **By-subjects (F1)**: Average RTs per condition for each participant; participant is the random factor
+2. **By-items (F2)**: Average RTs per condition for each item; item is the random factor
+
+Both F1 and F2 must be significant for the effect to be considered reliable (Clark, 1973).
+
+**Limitations**: By-subjects + by-items ANOVA is less powerful than LMMs, cannot handle unbalanced data (e.g., from trial exclusion), and does not generalize simultaneously over both subjects and items (Baayen et al., 2008). Use LMMs when possible.
+
+### 3.3 Bayesian Mixed-Effects Models
+
+For studies where null effects are theoretically informative:
+
+```r
+library(brms)
+
+model <- brm(log_RT ~ condition +
+ (1 + condition | subject) +
+ (1 + condition | item),
+ data = data,
+ family = lognormal())
+```
+
+**Advantages**: Provides evidence for null effects (via Bayes factors or posterior distributions), handles convergence issues better than frequentist LMMs, and naturally handles the lognormal distribution of RTs.
+
+**Reference**: Nicenboim & Vasishth (2016); Vasishth, Mertzen, Jager, & Rabe (2019).
+
+---
+
+## 4. Residual Reading Time Analysis
+
+Residual reading times (Ferreira & Clifton, 1986; Trueswell, Tanenhaus, & Garnsey, 1994) remove variance due to word length (and optionally frequency), isolating variance due to the experimental manipulation.
+
+### Procedure
+
+1. For each participant, fit a simple linear regression predicting RT from word length (in characters) across all filler items:
+
+ ```r
+ participant_model <- lm(RT ~ nchar, data = fillers_for_this_participant)
+ ```
+
+2. Use this regression to predict expected RT for each word in the experimental items:
+
+ ```r
+ predicted_RT <- predict(participant_model, newdata = experimental_items)
+ ```
+
+3. Compute residual RT:
+
+ ```r
+ residual_RT <- observed_RT - predicted_RT
+ ```
+
+4. Analyze residual RTs as the dependent variable.
+
+### When to Use
+
+- When critical regions differ in word length across conditions and matching is impossible
+- As a robustness check alongside raw or log RT analyses
+- When you want to ensure that your effect is not driven by word length differences
+
+### When NOT to Use
+
+- When word length is perfectly matched across conditions (residualization is unnecessary)
+- When using LMMs with word length as a covariate (this achieves the same goal within the model)
+- As the sole analysis -- always report at least one analysis on the raw or log-transformed scale alongside residual RTs
+
+---
+
+## 5. Spillover Region Analysis
+
+Spillover is the hallmark of SPR data. Effects at a critical word frequently appear on subsequent words, not the critical word itself (Mitchell, 2004; Rayner, 1998).
+
+### Which Regions to Analyze
+
+| Region | Label | Typical Pattern |
+|---|---|---|
+| Critical word (CW) | CW or R0 | Effect may or may not appear here |
+| CW + 1 | Spillover 1 | Often the largest effect in SPR |
+| CW + 2 | Spillover 2 | Effect may still be present |
+| CW + 3 | Spillover 3 | Rarely significant, but worth checking |
+| Sentence-final word | Wrap-up | Inflated by wrap-up; do NOT treat as spillover |
+
+### Analysis Approaches
+
+#### Approach A: Region-by-Region Analysis
+
+Run a separate statistical model for each region (CW, CW+1, CW+2). This is the most common approach.
+
+**Advantage**: Simple, easy to interpret.
+**Disadvantage**: Inflates familywise error rate due to multiple comparisons. Apply correction (e.g., report whether effects survive Bonferroni or Holm correction across regions).
+
+#### Approach B: Region as a Factor
+
+Include region (CW, CW+1, CW+2) as a factor in a single model:
+
+```r
+model <- lmer(log_RT ~ condition * region +
+ (1 + condition * region | subject) +
+ (1 + condition * region | item),
+ data = data)
+```
+
+**Advantage**: Controls familywise error; the condition x region interaction directly tests whether the effect changes across regions.
+**Disadvantage**: Model complexity; may not converge with full random effects.
+
+#### Approach C: Including Previous-Region RT as a Covariate
+
+Include the RT from the previous region as a predictor to control for spillover autocorrelation:
+
+```r
+model <- lmer(log_RT ~ condition + scale(prev_RT) +
+ (1 + condition | subject) +
+ (1 + condition | item),
+ data = data)
+```
+
+**Advantage**: Partials out autocorrelation; a significant condition effect in this model indicates processing difficulty above and beyond what is explained by the previous word.
+**Disadvantage**: Removes some variance that is part of the effect of interest; use with caution and always report alongside the model without this covariate.
+
+---
+
+## 6. Handling Missing Data
+
+### Sources of Missing Data in SPR
+
+| Source | Typical Rate | How to Handle |
+|---|---|---|
+| **Timeouts** (if using a response deadline) | 1-5% | Treat as missing; do not substitute an arbitrary value. Report the timeout rate per condition -- if it differs between conditions, the timeout rate itself may be informative |
+| **Incorrect comprehension trials** | 5-15% | See Section 1.2 above |
+| **RT trimming** (absolute + SD cutoffs) | 2-5% | Already handled by the trimming pipeline |
+| **Technical errors** (software crashes, key misregistration) | <1% | Exclude those trials |
+
+### What NOT to Do
+
+- **Do not replace missing RTs with the condition mean.** This artificially reduces variance and inflates power.
+- **Do not replace missing RTs with the participant mean.** Same problem.
+- **Do not use listwise deletion** (removing the entire sentence if one word is missing). This is wasteful. LMMs handle missing data naturally by estimating from available observations.
+
+### What to Do
+
+- Use **linear mixed-effects models**, which handle unbalanced data and missing observations without requiring imputation (Baayen et al., 2008).
+- **Report** the amount and pattern of missing data. If missing data is concentrated in one condition (e.g., more timeouts in the harder condition), this is informative and should be discussed.
+
+---
+
+## 7. Reporting Standards
+
+### What to Report in an SPR Analysis
+
+| Element | Details |
+|---|---|
+| **RT trimming criteria** | Lower cutoff (e.g., 100 ms), upper cutoff (e.g., 2500 ms), and any SD-based trimming |
+| **Percentage of data excluded** | At each trimming step and in total |
+| **Number of participants excluded** | With reason (accuracy threshold, technical issues) |
+| **RT transformation** | Log, inverse, or none; justify the choice |
+| **Model specification** | Full model formula including random effects structure |
+| **Convergence** | Whether the model converged; if simplification was needed, what was removed |
+| **Results by region** | Report statistics for the critical region AND at least 2 spillover regions |
+| **Effect direction** | Report mean RTs (not just test statistics) for each condition at each region |
+| **Descriptive statistics** | Mean RT per condition per region with standard errors, ideally plotted |
+| **Comprehension accuracy** | Overall and by condition |
+
+### Visualization
+
+Plot mean reading times by region and condition. The standard figure is a **region-by-region plot** with:
+- X-axis: word position or region label
+- Y-axis: mean reading time (ms) -- use raw ms even if the analysis used log RT, for interpretability
+- Separate lines for each condition
+- Error bars: standard error of the mean (computed across participants)
+
+This plot allows readers to see the time course of the effect, including whether it appears at the critical word, spills over, or is delayed.
+
+---
+
+## 8. Effect Size Benchmarks
+
+SPR effects are typically smaller than effects in speeded choice tasks (e.g., Stroop, lexical decision). Calibrate your expectations:
+
+| Effect Type | Typical Effect Size (ms) | Example |
+|---|---|---|
+| **Garden-path effect** (strong) | **40-100 ms** at disambiguation point | Reduced relative clause (Ferreira & Clifton, 1986) |
+| **Plausibility mismatch** | **30-80 ms** | Implausible agent/patient (Garnsey et al., 1997) |
+| **Agreement violation** | **20-60 ms**, often mostly in spillover | Number attraction (Pearlmutter et al., 1999) |
+| **Frequency effect** | **20-40 ms per log unit** | High vs. low frequency words (Rayner, 1998) |
+| **Subtle pragmatic effect** | **10-30 ms** | Scalar implicature, presupposition (Bott & Noveck, 2004) |
+
+These benchmarks should inform power analysis. For a typical 30 ms effect with within-participant SD of ~200 ms, you need approximately **24-32 items per condition** and **40+ participants** for 80% power in a mixed-effects model (Brysbaert & Stevens, 2018).
+
+---
+
+## References
+
+- Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. *Journal of Memory and Language, 59*, 390-412.
+- Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing. *Journal of Memory and Language, 68*, 255-278.
+- Bates, D., Kliegl, R., Vasishth, S., & Baayen, H. (2015). Parsimonious mixed models. arXiv:1506.04967.
+- Bott, L., & Noveck, I. A. (2004). Some utterances are underinformative. *Journal of Memory and Language, 51*, 437-457.
+- Box, G. E. P., & Cox, D. R. (1964). An analysis of transformations. *Journal of the Royal Statistical Society, Series B, 26*, 211-252.
+- Brysbaert, M., & Stevens, M. (2018). Power analysis and effect size in mixed effects models. *Journal of Cognition, 1*(1), 9.
+- Clark, H. H. (1973). The language-as-fixed-effect fallacy. *Journal of Verbal Learning and Verbal Behavior, 12*, 335-359.
+- Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. *Journal of Memory and Language, 25*, 348-368.
+- Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. *Journal of Memory and Language, 37*, 58-93.
+- Jegerski, J. (2014). Self-paced reading. In J. Jegerski & B. VanPatten (Eds.), *Research methods in second language psycholinguistics*. Routledge.
+- Lo, S., & Andrews, S. (2015). To transform or not to transform: Using generalized linear mixed models to analyse reaction time data. *Frontiers in Psychology, 6*, 1171.
+- Marsden, E., Thompson, S., & Plonsky, L. (2018). A methodological synthesis of self-paced reading in second language research. *Applied Psycholinguistics, 39*, 861-904.
+- Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., & Bates, D. (2017). Balancing Type I error and power in linear mixed models. *Journal of Memory and Language, 94*, 305-315.
+- Mitchell, D. C. (2004). On-line methods in language processing. In M. Carreiras & C. Clifton (Eds.), *The on-line study of sentence comprehension*. Psychology Press.
+- Nicenboim, B., & Vasishth, S. (2016). Statistical methods for linguistic research. *Language, Cognition and Neuroscience, 31*, 748-767.
+- Pearlmutter, N. J., Garnsey, S. M., & Bock, K. (1999). Agreement processes in sentence comprehension. *Journal of Memory and Language, 41*, 427-456.
+- Ratcliff, R. (1993). Methods for dealing with reaction time outliers. *Psychological Bulletin, 114*, 510-532.
+- Rayner, K. (1998). Eye movements in reading and information processing. *Psychological Bulletin, 124*, 372-422.
+- Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on parsing. *Journal of Memory and Language, 33*, 285-318.
+- Vasishth, S., Mertzen, D., Jager, L. A., & Rabe, A. (2019). The statistical significance filter leads to overoptimistic expectations of replicability. *Journal of Memory and Language, 108*, 104027.
+- Vasishth, S., & Nicenboim, B. (2016). Statistical methods for linguistic research: Foundational ideas - Part I. *Language and Linguistics Compass, 10*, 349-369.