What do drawing and speaking capture about how children express their knowledge of the world?

How much overlap exists between children's visual and linguistic expressions of the same concept, and how does this overlap shift across development? This repository contains the analysis scripts and data for a study in which children (ages 2–12) drew and described six sea animal categories (crab, octopus, seahorse, shark, turtle, whale) at the Birch Aquarium.

To reproduce all analyses and figures, you only need data/tidy/ and the scripts in analysis/. The preprocessing scripts (preprocessing/01–05) are included for transparency and document how the tidy data was generated from raw drawings and transcripts, but are not runnable without access to the raw data.

---

Participants

• Children: BD025–BD124 (BD001–BD024 are pilot participants, excluded)
• Adults: AD001–AD020 (used to compute adult prototype embeddings)
• BD085 and BD097 are excluded from age-based analyses due to a data collection error (missing age)

---

Repository structure

data/
  raw/               # Raw MongoDB output (drawings, metadata)
  clean/             # Cleaned metadata, transcripts, age table
  embeddings/        # CLIP embeddings for drawings and utterances
  preprocessed/      # Cosine similarity outputs (pre-tidy)
  tidy/              # Canonical analysis-ready CSVs (start here)
    drawings.csv
    utterances.csv
data_intermediates/
  summaries/         # trial_summaries.txt — participant/trial counts
preprocessing/       # Scripts 01–05
analysis/            # Scripts 06–09b
figures/             # All output figures

---

Tidy data

The two files in data/tidy/ are the canonical starting point for all analyses. Each has one row per participant × sea animal category and includes adults and children, with is_adult/is_child flags.

drawings.csv — drawing metrics per participant × category:

• draw_exists, draw_interference — exclusion flags
• trial_duration, num_strokes, draw_duration, mean_intensity, bounding_box — drawing metadata
• draw_cossim_adults, draw_half_cossim_adults — cosine similarity to adult prototype at 100% and 50% of drawing
• draw_cossim_{category}, draw_probability_{category} — within-subject similarity for RSM analyses
• draw_full_features, draw_half_features — hand-coded visual features at 100% and 50%

utterances.csv — description metrics per participant × category:

• utterance_exists, utterance_interference — exclusion flags
• utterance_ns, utterance_masked — raw and name-scrubbed transcripts
• utterance_full_masked_filtered, utterance_half_masked_filtered — filtered transcripts
• utterance_full_cossim_adults, utterance_half_cossim_adults — cosine similarity to adult prototype at 100% and 50% of utterance
• utterance_cossim_{category}, utterance_probability_{category} — within-subject similarity for RSM analyses
• utterance_full_visual_features, utterance_full_semantic_features — hand-coded features at 100%
• utterance_half_visual_features, utterance_half_semantic_features — hand-coded features at 50%

Standard exclusion filters used in all analysis scripts:

drawings <- read_csv("data/tidy/drawings.csv") %>%
  filter(is_child, draw_exists == 1, draw_interference == 0)

utterances <- read_csv("data/tidy/utterances.csv") %>%
  filter(is_child, utterance_exists == 1, utterance_interference == 0)

---

Preprocessing pipeline

Scripts 01–05 are in preprocessing/ and are included for documentation only — they are not runnable without the raw MongoDB data and drawing image files. Run scripts in order if you have access to the raw data.

01_draw_mongo_retrieval

Extracts drawings and stroke data from the MongoDB database.

	Path
Output	data/raw/mongo_output/highres/sketched_full_dataset/
Output	data/raw/mongo_output/highres/AllDescriptives_images_final_birch_run_v1.csv

02_clean_data

Excludes pilot participants (BD001–BD024), adds age data, renames columns to the format the embedding generator expects (image1 for file path, text1 for category).

	Path
Input	data/raw/mongo_output/highres/AllDescriptives_images_final_birch_run_v1.csv
Input	data/clean/participant_age_table.csv
Input	data/clean/half_and_full_utterance_features.csv
Output	data/clean/drawing_highres_clean.csv
Output	data/clean/transcripts_raw_clean.csv

03_draw_generate_clip_embedding

Generates CLIP image embeddings for all drawings. Note: due to a known issue with the embedding generator's output_path argument, embeddings are written to preprocessing/output/image_embeddings/ and must be moved by hand to data/embeddings/highres_sketch_embeddings/image_embeddings/ before running the second half of the script. The second half column-wise z-scores the embeddings, merges metadata, and filters to sea animal categories only.

	Path
Input	data/clean/drawing_highres_clean.csv
Output (hand-moved)	data/embeddings/highres_sketch_embeddings/image_embeddings/clip_image_embeddings_csv.csv
Output	data/embeddings/highres_sketch_embeddings/image_embeddings/draw_embeddings_with_metadata.csv

03_talk_generate_clip_embedding

Filters utterances, generates CLIP text embeddings from the masked utterance column, and column-wise z-scores them.

	Path
Input	data/clean/transcripts_raw_clean.csv
Output	data/embeddings/utterance_half_and_full_embeddings.csv

04a_draw_generate_cossim_adults

Computes adult prototype embeddings (mean over adult participants) per category, then computes each child's cosine similarity to that prototype for both the full drawing and the 50%-timepoint drawing. Half-time embeddings are column-wise z-scored here. Adults are dropped from the output.

	Path
Input	data/embeddings/highres_sketch_embeddings/image_embeddings/draw_embeddings_with_metadata.csv
Input	data/embeddings/half_time_sketch_embeddings/
Output	data/preprocessed/draw_cossim_adults.csv

04b_talk_generate_cossim_adults

Same as 04a but for utterances. Produces utterance_full_cossim_adults and utterance_half_cossim_adults.

	Path
Input	data/embeddings/utterance_half_and_full_embeddings.csv
Output	data/preprocessed/talk_cossim_adults.csv

04c_draw_generate_cossim_toself

For each participant × category drawing, computes cosine similarity to the participant's own drawings of the other five categories. Applies softmax to yield draw_probability_{category}. Adults included.

	Path
Input	data/embeddings/highres_sketch_embeddings/image_embeddings/draw_embeddings_with_metadata.csv
Output	data/preprocessed/draw_cossim_toself.csv

04d_talk_generate_cossim_toself

Same as 04c but for utterances. Yields utterance_probability_{category}. Adults included.

	Path
Input	data/embeddings/utterance_half_and_full_embeddings.csv
Output	data/preprocessed/talk_cossim_toself.csv

05_create_tidy_data

Combines all metadata, hand-coded features, and preprocessed cosine similarity values into the two canonical tidy CSVs. Excludes participants with missing age data (BD085, BD097).

	Path
Input	data/clean/half_and_full_utterance_features.csv
Input	data/clean/participant_age_table.csv
Input	data/clean/drawing_highres_clean.csv
Input	data/embeddings/half_time_sketch_embeddings/half_time_clip_embeddings.csv
Input	data/preprocessed/draw_cossim_adults.csv
Input	data/preprocessed/draw_cossim_toself.csv
Input	data/preprocessed/talk_cossim_adults.csv
Input	data/preprocessed/talk_cossim_toself.csv
Output	data/tidy/drawings.csv
Output	data/tidy/utterances.csv

---

Analysis scripts

All scripts are in analysis/ and read from data/tidy/. Run independently of one another (no ordering requirement).

06_correlation_analysis

Computes within-modality age effects and cross-modal correlation between drawing and description similarity to adults. Produces Figs 3A, 3B, 4.

07_rsm_toself_analysis

Writes participant/trial summary statistics, then builds representational similarity matrices (RSMs) from within-subject softmax probabilities, correlates RSMs across modalities for younger vs. older children, and plots cross-modal scatter. Produces Figs 5, 6.

08_features_analysis

Plots hand-coded feature proportions by category and feature counts by age, for drawings and descriptions at both 50% and 100% timepoints. Produces Figs 7, 8, 9, 10, 13, 14, 15, 16.

09a_draw_time_analysis

Plots drawing cosine similarity to adult prototype at 50% vs. 100% of drawing (z-scored within category), fits mixed models for timepoint and age effects, and plots developmental trajectories by age bin. Produces Figs 11A, 12A.

09b_talk_time_analysis

Same as 09a but for utterances. Produces Figs 11B, 12B.