Class results
This page is not connected to the class data yet. Paste the Apps Script web-app URL into assets/config.js (see README).
Summary
Translation priming, one dot per person
Reaction times
Correct responses to real words only, 200–5,000 ms, as in Chaouch-Orozco et al. (2021, 2023). Means of each person's own mean. Test: each person's priming effect against 0 (a one-sample t-test, the same as a paired t-test of translation vs unrelated); dz = mean effect ÷ its SD. Two-tailed, α = .05.
Accuracy
Every word trial the student saw properly, right or wrong, with no reaction-time trimming. Priming = errors after unrelated − errors after translation, in percentage points; positive = more errors without the translation. Same test as for reaction times. Non-words are shown as a check that students were doing the task.
Did people see the masked primes?
Whose screen showed what
Summary
Round-shape choices by sound
Cell means
Tests
Paired t-tests on each student's own percentages (comparisons), and one-sample t-tests against 50% (chance). Two-tailed, α = .05.
The data script has not been updated for this task yet, so its results cannot be stored or shown. Update Code.gs and deploy a new version (see README).
Summary
Gambles accepted
Continuing after a sunk cost
Superstition
English proficiency and the effect
Each dot is a student in the English version. Black line: regression on English self-rating, with its 95% band. Dashed line: the Chinese group's mean. If the effect fades as English improves, the line moves toward the dashed line at higher ratings.
Slope = change in the score per 1-point rise in the self-rating, tested against 0. Interaction = English slope − Chinese slope, with a 95% CI and its test; a significant interaction means proficiency changes the effect. In the Chinese version, English proficiency should not matter.
All measures
Means of each student's own score. Difference = English − Chinese, with a 95% confidence interval. Welch's t-test (unequal variances), two-tailed, α = .05; d = difference ÷ pooled SD. Only students with Chinese as a first language who used no translation aid.
English self-ratings
1 = very limited, 7 = native-like. The two groups should be similar (Welch's t-test).
Each student gave chains of ten associations from twenty starting words, ten in Chinese and ten in English. Two questions: what happens from one answer to the next (transition probabilities: the chance that the next answer is positive or negative, given the one before), and how a chain develops over its ten answers (trajectories: how average valence changes from answer 1 to 10).
What we found
Updating the models with the newest runs…
ADo transition probabilities differ between Chinese and English?
Each answer counts as negative (valence below 5 on the 1–9 scale) or positive. A transition probability is the chance that the next answer is negative or positive, given the answer before it.
Observed transition probabilities
These are the raw proportions in the class data; the statistical comparisons below come from the mixed-effects model. Each row sums to 1.
Model-adjusted transition probabilities
Continuous-valence check: slopes in Chinese and English and their difference, still being fitted
Instead of dividing answers into positive and negative states, this analysis asks how strongly the exact valence of one answer predicts the valence of the next.
Three-state transition dynamics: still being fitted
The two-state analysis asks whether an answer is negative or positive. This additional analysis separates a neutral middle state, allowing us to ask whether Chinese and English differ in movements towards negative, neutral or positive conceptual states. Negative: below 4.2; neutral: 4.2 to below 5.8; positive: 5.8 and above (English scale). It uses the same transitions as the two-state analysis, with a multinomial mixed-effects model (see Methods).
Observed three-state transition probabilities
Raw proportions in the class data, with counts. Rows are the previous answer and sum to 1. The step from the starting word to the first answer is not counted.
Model-adjusted three-state transition probabilities
Check without cut-points: next valence as a curve of previous valence
The same transitions on the continuous scale: the next answer’s valence modelled as a curve (linear and squared terms) of the previous answer’s valence, in each language. If the languages differ in where answers move from each part of the scale, the two curves differ in shape; this needs no negative, neutral and positive boundaries.
Numbers, and the share of answers in each state
Other cut-points (sensitivity)
The 4.2 / 5.8 cut-points are an analytical choice. The same model with two other neutral bands, fixed in advance (not searched for): 4.0 / 6.0, and a narrower band, 4.5 / 5.5.
Leave one starting word out (sensitivity): fitted when opened
The three-state model refitted 20 times, each time without one of the 20 starting words (10 per language), to see whether any single word drives the result. Only the range across the refits is reported; the refits are not separate findings. They are fitted when this section is opened and take a few minutes.
Additional transition measure
Counts behind the observed probabilities
Rows are the previous answer and sum to 1. The step from the starting word to the first answer is not counted. The three-state counts are under “Three-state transition dynamics”.
BDo valence trajectories differ between Chinese and English?
A trajectory is how the average valence of answers changes from answer 1 to answer 10.
Does the starting word influence the trajectory differently by language?
Every chain begins from a starting word (the “seed”) that is negative, neutral or positive. These words first push the chain towards different parts of the valence scale. This asks how long that influence stays visible, and whether it fades differently in Chinese and English.
Observed trajectories after each kind of starting word
Observed means, not model estimates. These plots show how answer valence develops across the ten answers after differently valenced starting words. Bands: 95% confidence intervals across students. Intervals appear from 3 students.
Numbers by starting word
Observed. Staying negative and staying positive: pooled transitions. Valence: each student’s mean over scored answers, then the class mean (1–9). Repeated answers: the share of answers that repeat an earlier word of the same chain, the starting word included.
COther behavioural differences
English − Chinese, with 95% confidence intervals; the vertical line is 0, no difference.
Response timing
Timing with and without extreme medians
Repetition
Show individual data
Response time
Onset (first keystroke)
Total (answer submitted)
Observed. Each student’s median time per answer, in seconds, on a log scale (the scale the timing models use); every student is shown at their own value. Grey lines join each student’s two medians; the marks at the edges are geometric means with 95% confidence intervals, comparable with the models’ predicted times.
Repeated answers
Exploratory: Does English proficiency change the language difference?
An exploratory moderation analysis: does the English − Chinese difference change with English proficiency (self-rated, 0 = no knowledge to 10 = native-like)? The transition and valence models are refitted with each student’s rating and its interactions with language, with by-student random slopes for the language terms. Measures: staying negative, staying positive and mean answer valence; the overall rating is primary.
Show the plots
Skill-specific follow-ups
The same models for listening, speaking, reading and writing. These twelve tests are judged on Benjamini–Hochberg adjusted p-values; they are fitted after the overall rating, so they appear a little later.
Change in the English − Chinese difference for each extra rating point (log-odds for staying negative and positive, valence points for mean valence), with 95% confidence intervals; 0 = no change with proficiency. Filled: clear after adjustment. Overall* is the primary rating, shown for comparison.
Change per point: from the proficiency models (see Methods). Staying negative and staying positive: log-odds, Wald z; mean valence: valence points, Satterthwaite t. The overall rating is the primary test for each measure (three tests, unadjusted).
Exploratory: What might produce the trajectory difference?
These analyses were motivated by patterns noticed after inspecting the class results and are exploratory. They should be treated as hypotheses for future work rather than confirmatory evidence.
Show the exploratory analyses
1 · Do English chains converge on a smaller response vocabulary?
For each student and language: how often a response recurs across the student’s own different chains (cross-chain reuse), the share of distinct responses, and how often an answer repeats its chain’s starting word or an earlier answer in the same chain (within-chain repetition). This is reuse within one person’s chains, not a vocabulary shared across students.
Definitions and the full table
2 · Does lexical convergence go with the starting-word decay difference?
The starting-word model again, with every term also multiplied by each student’s English − Chinese difference in cross-chain reuse (centred), and a by-student random slope for the decay term. The question: do students whose English chains reuse more across chains (relative to their Chinese chains) also show a larger English − Chinese difference in how fast the starting-word gap shrinks?
3 · Are L1 (Chinese) dynamics associated with bilingual experience?
Chinese chains only, with three measures of bilingual experience: English age of acquisition, current English use, and the English − Chinese overall self-rating. Each is tested separately against three features of the Chinese chains, a family of nine tests judged on Benjamini–Hochberg adjusted p-values. These data cannot distinguish bilingualism-induced L1 change from pre-existing individual differences because there is no monolingual comparison group.
4 · Semantic-space analysis (not done)
Exact-response reuse is not semantic convergence. Testing whether English chains move away from the starting concept faster, or explore less of the semantic space, needs a validated multilingual embedding model that places Chinese and English words and short phrases in one space, applied with the same preprocessing to both languages and run offline into a fixed table of every response (updated when a class adds new responses). The page has no such model, and running one in the browser would make it slow, so this is left for future work.
Robustness and data quality
Limitation: Chinese and English used separate matched sets of starting words, not translations, so the language comparisons cannot completely separate language from differences between the two word sets.
Robustness checks
Which answers could not be scored?
The column “By-student random slopes” adds uncorrelated by-student slopes for the language terms. materials/fast/random-slopes.R fits the transition model with correlated slopes in R, from the Transitions download.
Coverage threshold (sensitivity)
The main analysis keeps every student who passes the repetition and first-language rules, however many of their answers the norms could score. Here the primary models and the three-state model are refitted on only the students with at least 50%, 60%, 70% or 80% of their possible transitions usable in both languages. This shows how much the conclusions depend on including students with low coverage.
Participant-level data checks
Counts only. The exclusion criteria are stated at the top of the page; the robustness column “All students (no exclusions)” shows the main models without them.
Data quality
Answers scored: found in the valence norms (Warriner et al., 2013; Chan & Tse, 2024). Valid transitions: pairs of adjacent answers that were both scored, out of 90 per student and language (answers 1→2 to 9→10 in ten chains). An unscored answer removes the transitions on either side of it; transitions are never bridged across a missing valence rating.
Methods and full statistical output
Methods and model specifications
Scoring
Each answer is looked up in the valence norms: Warriner et al. (2013) for English (after lower-casing, British → American spelling and lemmatising), Chan & Tse (2024) for Chinese (simplified or traditional form). Answers the norms do not rate stay in their chain as missing.
The Chinese norms are more compressed around the middle of the scale, so they are linearly mapped onto the English normative scale before cross-language comparison. The map gives Chinese ratings the English mean and standard deviation over the two whole norm sets: English-scale valence = −1.53 + 1.31 × Chinese valence (assets/data/equating.json). The percentile-equating check in Robustness uses a non-linear map instead.
States: negative below 5, positive 5 and above (on the English scale). Transitions are pairs of adjacent scored answers, from answer 1→2 to 9→10; the starting word → answer 1 step is not counted.
Coding
Language: English = +½, Chinese = −½. Previous state: positive = +½, negative = −½. Block: second language done = +½, first = −½. Seed valence: the starting word’s valence on the English scale, minus its mean over the rows the model uses. Every model has random intercepts for students, starting words (seeds) and chains (one student’s ten answers to one starting word).
Design limitation: the starting words
Chinese and English used separate matched sets of starting words rather than translation equivalents. Language comparisons therefore generalise across these matched sets, but cannot completely separate language from item-set differences. The random intercept for starting words absorbs variation among the words; it does not turn the comparison into an item-matched one.
Transition model (logistic)
logit P(next answer positive) = β0 + β1·Language + β2·Previous state + β3·Language × Previous state + β4·Seed valence + β5·Position + β6·Block + ustudent + useed + uchain
One row per valid transition. Position: the earlier answer’s position minus 5 (transitions 1→2 to 9→10). Laplace approximation, with β estimated jointly with the random effects (as lme4’s nAGQ = 0).
Continuous-valence check (linear)
Next valence = β0 + β1·Language + β2·(Previous valence − 5) + β3·Language × (Previous valence − 5) + β4·Seed valence + β5·Position + β6·Block + ustudent + useed + uchain + ε
Trajectory model (linear)
Valence = β0 + β1·Language + β2·P + β3·(P² − 8.25) + β4·Language × P + β5·Language × (P² − 8.25) + β6·Seed valence + β7·Block + ustudent + useed + uchain + ε
P = position − 5.5. One row per scored answer (positions 1–10). Because P is centred at the middle of the chain and the model also has a curve, β4 is the language difference in slope at the centre of the chain, not a constant difference from position 1 to 10. The sensitivity check replaces P and P² with position as a category (answer 1 the reference).
Starting-word model (linear)
Valence = β0 + β1·Language + β2·Seed valence + β3·log2(position) + β4·Language × Seed valence + β5·Language × log2(position) + β6·Seed valence × log2(position) + β7·Language × Seed valence × log2(position) + β8·Block + ustudent + useed + uchain + ε
log2(position) is 0 at answer 1, so β4 is the language difference in the starting word’s influence on the first answer, and β7 the language difference in how that influence changes each time the position doubles. The separation plot multiplies the model’s influence per point of seed valence by the gap between the mean valence of all positive and all negative starting words (both languages’ words together, the same value for both languages).
Proficiency models (exploratory)
logit P(next answer positive) = [the transition model’s terms] + β7·R + β8·Language × R + β9·Previous state × R + β10·Language × Previous state × R + ustudent + ustudent,L·Language + ustudent,LP·(Language × Previous state) + useed + uchain
Valence = [the trajectory model’s terms] + β8·R + β9·Language × R + ustudent + ustudent,L·Language + useed + uchain + ε
R = the student’s English self-rating (0–10) minus the mean over the students in the model; one pair of models per rating. The rating varies only between students, so the language terms get by-student random slopes (uncorrelated, as lme4’s ||); without them the standard errors of these cross-level interactions would be too small. Change in the English − Chinese difference per rating point: staying positive β8 + ½β10, staying negative −β8 + ½β10 (log-odds, Wald z); mean valence β9 (valence points, Satterthwaite t). The plots’ lines are the models’ English − Chinese differences at each rating (average marginal predictions for the transition model). Every student’s transitions count, weighted by how many they have, rather than one proportion per student. The overall rating is primary (three tests, unadjusted); the four skills’ twelve tests are judged on Benjamini–Hochberg adjusted p-values. This is an exploratory moderation analysis: ratings are self-reports, and the class was not sampled to vary in proficiency.
Three-state transition model (multinomial)
log [P(next = k) / P(next = negative)] = β0k + β1k·Language + β2k·Prev neutral + β3k·Prev positive + β4k·Language × Prev neutral + β5k·Language × Prev positive + β6k·Seed valence + β7k·Position + β8k·Block + ustudent,k + useed,k + uchain,k, k = neutral, positive
States on the English scale: negative below 4.2, neutral 4.2 to below 5.8, positive 5.8 and above (4.2 is neutral; 5.8 is positive). One row per valid transition: the same rows as the two-state model (adjacent answers both scored, 1→2 to 9→10; nothing is bridged across an unscored answer). Outcome: the next answer’s state, a baseline-category (multinomial) logit with negative as the reference destination, so each destination has its own set of coefficients. Previous state: indicators for neutral and positive (negative the reference), so β1k is the language difference after a negative answer; Language, Block, Seed valence and Position are coded as in the two-state model. Random intercepts for students, starting words and chains, each a pair (one per logit) with an unstructured 2 × 2 covariance, so the fit does not depend on which destination is the reference. Laplace approximation with β estimated jointly with the random effects (as lme4’s nAGQ = 0, like the two-state model); the covariance parameters by Nelder–Mead. Checked in R: an independent implementation reproduces the fit, tests and probabilities; RTMB’s full Laplace fit (β outside the integral, as nAGQ = 1) gives probabilities within .011 and the same verdicts for the omnibus test, the Holm-adjusted tests and the two named contrasts.
Omnibus test: joint Wald χ²(6) that the six language terms (β1, β4, β5 for both logits) are zero, i.e. that given the previous state and the covariates the destination probabilities do not differ by language. Follow-ups, after each previous state: a joint χ²(2) of the language difference in both logits at that state (β1, β1 + β4 or β1 + β5), Holm-adjusted across the three. Model-adjusted probabilities: average marginal predictions as for the two-state model (every observed transition predicted with language and previous state set, random effects at zero, then averaged; each row sums to 1, which the page checks); differences with delta-method CIs. Two contrasts were named in advance, negative → positive and positive → positive; they are read only together with their row’s test. Robustness: all students, no chain intercepts, by-student language slopes (a 2 × 2 covariance, uncorrelated with the intercepts), percentile equating. Cut-point sensitivity: 4.0 / 6.0 and 4.5 / 5.5. Leave-one-starting-word-out: the model refitted without each of the 20 starting words in turn; only the range of the omnibus test and of the two named contrasts is reported, and how many refits keep the main verdicts.
Check without cut-points (linear mixed model)
Next valence = β0 + β1·Language + β2·V + β3·V² + β4·Language × V + β5·Language × V² + β6·Seed valence + β7·Position + β8·Block + ustudent + useed + uchain + ε, V = previous valence − 5
The same transitions as the two- and three-state models. Test: the joint F-test (Satterthwaite) of β4 and β5, whether the next answer’s valence depends on the previous answer’s differently in the two languages. It is the three-state model’s robustness check without cut-points: if it is not clear while the three-state verdicts change with the cut-points, the page treats the three-state evidence as threshold-sensitive, descriptive and secondary. Checked against lmerTest.
Response timing
log(median time) = β0 + β1·Language + β2·Block + β3·Language × Block + ustudent + ε
Fitted separately for onset (answer shown → first keystroke) and total time (answer shown → answer submitted), on each student’s median per language: two rows per student. Block: second language done = +½, first = −½, from the order the student did the languages in. With two rows per student, the language × block term is a between-student comparison: the two order groups differ by β3/2 in mean log time, shown as the ratio exp(β3/2), Chinese-first ÷ English-first. The block term is the practice effect averaged over the two languages; language-specific practice effects are not separable from the order groups here. Predicted times are exp(β0 ± β1/2), geometric means over students averaged over blocks; their difference in seconds has a delta-method CI; the other terms are shown as ratios, exp(β). Only these medians are stored with each student’s summary. Every answer’s onset and submit times are also kept in the data sheet (its fast_trials tab), but this page cannot read that tab, so no item-level timing model is fitted. The difference between a student’s median total time and median onset is not a median completion time and is not reported as one.
Timing outlier sensitivity
Rule, fixed in advance: on each student’s log median time, separately for Chinese onset, English onset, Chinese total and English total, robust z = 0.6745 × (x − median) / MAD (MAD: the median absolute deviation, unscaled), and |robust z| > 3.5 flags the value (where MAD is 0: outside Q1 − 3×IQR to Q3 + 3×IQR). A student is flagged for onset (or total) if either language’s value is. Flagged students are left out only of the matching timing model in the sensitivity analysis, which has the same specification as the primary one. They stay in the primary timing models and in every other analysis; being slow is not an exclusion criterion.
Lexical reuse (exploratory)
Three measures per student and language (defined under “What might produce the trajectory difference?”), each compared English − Chinese with a paired t-test, for all students and without the students with extreme repetition; Benjamini–Hochberg adjustment across the three measures within each sample. Sensitivity: raw, normalised and lemma response keys.
Exploratory models: reuse × starting-word decay, and L1 dynamics
Valence = [the starting-word model’s terms] + M × [each of them] + ustudent + ustudent,LSP·(Language × Seed valence × log2(position)) + useed + uchain + ε
M = the student’s English − Chinese cross-chain reuse (percentage points, normalised responses), centred. Key term: Language × Seed valence × log2(position) × M, reported in points of the gap’s change (answer 1 → 10) per percentage point of M.
Chinese only: Valence = β0 + β1·Seed valence + β2·log2(position) + β3·Seed valence × log2(position) + β4·Block + X × [β5 + β6·Seed valence + β7·log2(position) + β8·Seed valence × log2(position)] + ustudent + ustudent,SP·(Seed valence × log2(position)) + useed + uchain + ε
Chinese only: Valence = β0 + β1·P + β2·(P² − 8.25) + β3·Seed valence + β4·Block + X × [β5 + β6·P + β7·(P² − 8.25)] + ustudent + ustudent,P·P + useed + uchain + ε
X: English age of acquisition (years), English use (per 10 percentage points), or English − Chinese overall self-rating, each centred and tested in its own models. Key terms: β8 (starting-word persistence) and the joint test of β6 and β7 (trajectory shape). The third outcome, Chinese cross-chain reuse, is regressed on X by least squares with HC3 errors (one value per student). Benjamini–Hochberg adjustment across the nine tests. In every model with a student-level moderator, the term it moderates gets a by-student random slope (uncorrelated), so that the interaction’s standard error reflects differences between students.
Exclusion criteria
Applied before every analysis on the page (the CSV downloads keep every student, with columns analysed and excluded_for). Behavioural and data-quality rules only. A student is excluded when (1) more than half of their answers in a language repeat the chain’s starting word or an earlier answer in the same chain (the site’s repetition measure); or (2) their first-language answer here contradicts the same code’s answer to the Judgement task (“Is Chinese your first language?”), where both are loaded on the page. Normative coverage is not an exclusion criterion: it decides which answers can be scored, not whether an otherwise valid student is analysed. Unscored answers drop out as observations, with the transitions on either side of them. Response times are not trimmed or capped. The robustness column “All students (no exclusions)” refits the main models on everyone; “Coverage threshold” refits the primary models and the three-state model on only the students with at least 50%, 60%, 70% or 80% of their possible transitions usable in both languages. By-student random slopes (robustness): uncorrelated slopes for the language terms named in the robustness notes, added to each primary model.
Estimates, tests and intervals
Linear models: REML; t-tests and 95% CIs use Satterthwaite degrees of freedom (as lmerTest). Logistic model: Wald z and 95% CIs.
Model-adjusted probabilities (logistic model): average marginal predictions. Every observed transition is predicted with language and previous state set as stated, random effects at zero, keeping its own seed valence, position and block; the predictions are averaged. CIs by the delta method. Model-adjusted valence (linear models): the same averaging, which for a linear model equals the prediction at the observed means of the other predictors.
Omnibus tests are joint Wald tests that several coefficients are all zero. Transition pattern: Language and Language × Previous state (2 df, χ²). Trajectory shape: Language × P and Language × (P² − 8.25) (an F-test with Satterthwaite denominator degrees of freedom, as lmerTest’s contest). “Clear evidence” means a joint p < .05 for an omnibus test, and a 95% CI excluding zero for a single estimate or contrast. The page’s sentences follow the same rule: a sentence says two things differ, or that something rises, falls or shrinks, only when that difference or change has clear evidence; otherwise it gives the estimate and says it is not clear. No size word (“small”, “similar”, “much”) comes from a cut-off. Display-only rules, which change no conclusion: the note on observed vs model-adjusted probabilities appears when they differ at two decimals; students who repeated an earlier word in more than half of their answers are flagged under data quality; classes under 10 students get a “small class” note; CI bounds that round to zero get a third decimal. Every model and joint test was checked against lme4 and lmerTest in R on the class data, to the reported precision.
All tests
All terms of every model
Download the data
Summary
Typicality effect, one dot per person
Reaction times
True target statements only; the false fillers are never analysed by typicality. Correct responses, 250 ms ≤ RT < 5,000 ms (the time-out), trials with focus kept. Each person’s median per condition, then the class mean of those medians. Test: each person’s effect against 0 (one-sample t-test, the same as a paired t-test of low vs high); dz = mean effect ÷ its SD. Two-tailed, α = .05. A run counts if overall accuracy is at least 80% and at least 12 valid trials remain in each condition; only each person’s first such run counts.
Does English proficiency matter? exploratory
Before the task, students rated their English from 1 (very limited) to 7 (native-like) in five ways. Does the typicality effect change as the rating goes up?
All ratings
How much the typicality effect (low − high) changes with each extra rating point, with 95% confidence intervals. 0: no change with English proficiency. Bold and filled: significant.
One rating in detail
Fitted line, 95% band
Lower vs higher English
Typicality effect by English rating
Show the numbers
Slopes: least squares on each student’s English rating, counted runs only. Interaction: the slope of each student’s typicality effect, which equals the low-typicality slope minus the high-typicality slope. Six exploratory tests (five ratings and their mean), not corrected for multiple comparisons.
Summary
“Because” vs no connective
Each student read 10 critical passages with “because” and 10 without (lists A and B swap which). Recognition RT is the time to say that the first-clause verb (e.g. congratulated) had appeared.
Recognition RT
Second-clause reading time
Each student’s effect
Recognition RTs, trial by trial
Every counted trial (after the exclusions). Dashed line: the median trial; the table above gives the median of the students’ medians, so the two can differ slightly.
Across the session
Data quality
Exclusions only change what is counted here; the stored data are never changed, and every download keeps all observations (with the reason a trial was left out).
Students
Model-based analysis by students, by items, and mixed models of log RT
What this design can and cannot show
- “Because” also changes the punctuation and the sentence boundary: with it, the passage is one sentence; without it, two. A difference may come from either.
- Every critical probe is a word that did appear (and every filler probe one that did not), so the right answer can be guessed from the trial type, and false alarms on critical items cannot be measured.
- Faster recognition shows that the first-clause verb was easier to access; on its own it does not show that the two clauses were causally integrated.
- The early and final regions (first and last three words of the second clause) are approximate: the point where the causal link becomes clear differs between passages.
- Trial order, item differences and individual reading speed affect reading times; the counterbalanced lists and the by-student and by-item analyses address this only in part. This is a classroom study, not a confirmatory experiment.
Data
Import takes the trial and word CSV files (and the JSON file) saved by the class’s earlier HTML version of this task, or this page’s own downloads. Imported students stay in this browser only (they are not uploaded) and are marked I01, I02 …
Simulated data: 45 made-up students, generated in this browser for trying the dashboard. Nothing here is real or stored. Remove ?demo from the address for the real class.
Summary
Reaction times: same label vs different labels
Each student’s category advantage
How the class named the colours
For each colour, the share of students whose name was 青, 蓝, 绿 or another word both times, or who named it differently the two times (uncertain).
Where the names change
For each pair of neighbouring colours, the share of students whose (stable) names differ there — where the class draws its boundaries, and how much they vary.
Students
Exploratory model different labels, after accounting for colour distance
Cautions
- Colour naming differs between people: each student’s comparison rests on their own names, so students contribute different numbers of same- and different-label trials, and some cannot be compared at all.
- Screens are not calibrated: the same colour codes look different on different devices, brightness settings and lighting.
- Different-label pairs are not a random sample of pairs: they sit where names change, which may also be where colours are easier or harder to tell apart. The exploratory model accounts for the colour distance (two or three steps) and the particular pair, not for everything.
- Fast and wrong is possible: check accuracy alongside the reaction times (speed–accuracy trade-off).
- Few trials per student (60, fewer per condition): individual advantages are noisy.
- Naming the colours first may have made the names more active during matching.
- This is an observational classroom exercise: an association between names and speed cannot show that language causes a change in perception.
Data
The matching and combined files give each trial with the student’s own names for the two colours and its classification, so every result here can be recomputed. The stimuli’s colour values are in every file.
Summary
Prototypicality of birds
prcomp(scale. = TRUE) in Bird plot.Rmd). Larger, darker dot: higher mean typicality. Grey halo: disagreement, its radius proportional to the SD. The two axes are scaled separately. A component’s sign is arbitrary, so the map keeps the orientation of its first drawing for the whole class, also after a reload.Ratings by bird
1 = not at all typical, 5 = extremely typical. PC1 and PC2: the bird’s loadings, i.e. its place on the map.
Data
One row per student, with the columns of the Google Forms export, so Bird plot.Rmd runs on it unchanged (read it into responses1).
Summary
Punishment or reform?
Does the frame work the same in both languages?
All tests
English proficiency (exploratory)
Summary
The class’s mammal space
Do closer animals get stronger inferences?
Class map: each question’s mean answer
Own arrangements: every answer
Student by student
Does English proficiency matter? exploratory
At the end, students rated their English from 1 (very limited) to 7 (native-like) in five ways. Do students with stronger English show a closer link between their arrangement and their inferences?
All ratings
How much a student’s correlation (own arrangement) changes with each extra rating point, with 95% confidence intervals. 0: no change with English proficiency. Below 0: the link between closeness and inference is stronger with better English. Bold and filled: significant.
One rating in detail
Lower vs higher English
Correlation by English rating
English self-ratings
Show the numbers
Slopes: least squares of each student’s correlation (own arrangement) on their English rating; each student counts once (their first run). Six exploratory tests (five ratings and their mean), not corrected for multiple comparisons. Ratings are self-reports, and the class was not sampled to vary in proficiency.
All tests
Summary
Curiosity profiles
The class and the four profiles
Semantic networks
CC, ASPL and Q across the profiles
Students: does curiosity go with fluency?
Joyous exploration
Deprivation sensitivity
Combined (J + D)
Correlations
Both subscales together
The class’s answers
Animals
Fruits & vegetables
Entries that do not count
Animals
Fruits & vegetables
Each entry with the number of students who gave it and why it does not count under the chosen Count. Valid words are in SUBTLEX-CH and on the list for their block (animals; fruits and vegetables); the lists and their rules are in the repository (materials/fluency/categories).