library(tidyverse)
source("helpers.R")
library(ggrepel)
library(cowplot)
library(here)
library(colorspace)
library(ggeffects)Peekbank similarity analyses
Import data
usable_trials_summarized_with_sims <- read.csv(here("data/usable_trials_with_similarities.csv"))
aoa_ratings <- read.csv(here("data/metadata/level-aoaratings_type-kuperman_data.csv"))
saliency_values <- read.csv(here("data/metadata/level-imagepair_added-saliency_data.csv"))
usable_trials_summarized_with_sims <- usable_trials_summarized_with_sims |>
left_join(aoa_ratings |> transmute(text1=Word, aoa=AoA_Kup_lem)) |>
left_join(saliency_values, by=c("unique_pair"="ImagePair")) |>
mutate(age_bucket_bigger = cut(age,
breaks = c(12, 18, 24, 30, 36, 48, 60),
labels = c("(12-18]", "(18-24]", "(24-30]", "(30-36]", "(36-48]", "(48-60]"),
include.lowest = TRUE),
age_bucket = cut(age,
breaks = c(12, 24, 36, 48, 60),
labels = c("(12-24]", "(24-36]", "(36-48]", "(48-60]"),
include.lowest = TRUE))
# rounding each participant to the closest 5
age_based_trials <- usable_trials_summarized_with_sims |> mutate(
rounded_age = round_to_nearest(age, round_to=5)
)
clip_data_summarized <- summarize_similarity_data_collapsed(usable_trials_summarized_with_sims, extra_fields = c("dataset_name", "vanilla_trial", "aoa", "MeanSaliencyDiff")) |> mutate(
sim_logit = qlogis(pmin(pmax(image_similarity, 1e-6), 1 - 1e-6))
)
clip_data_summarized_age_bucket <- summarize_similarity_data_collapsed(usable_trials_summarized_with_sims, extra_fields = c("dataset_name", "vanilla_trial", "aoa", "MeanSaliencyDiff", "age_bucket")) |> mutate(
sim_logit = qlogis(pmin(pmax(image_similarity, 1e-6), 1 - 1e-6))
)
vanilla <- clip_data_summarized |> filter(vanilla_trial == 1)
sims <- c("image_similarity","text_similarity","multimodal_similarity","ooo_similarity")
sims_scaled <- c("scaled_image_similarity", "scaled_text_similarity", "scaled_multimodal_similarity", "scaled_ooo_similarity")
facet_labs <- c(text_similarity = "Text similarity",
image_similarity = "Image similarity",
multimodal_similarity = "Multimodal similarity",
ooo_similarity = "Adult behavioral similarity")
downloaded_aois <- readRDS(here("intermediates/downloaded_aois.Rds")) clip_data_summarized |> mutate(
sim_bucket = cut(
image_similarity,
breaks = seq(0, 1, by = 0.05),
include.lowest = TRUE,
right = FALSE
)
) |>
count(sim_bucket) |>
ggplot(aes(x = n, y = sim_bucket)) +
geom_col() +
labs(
x = "Count",
y = "Image similarity bucket",
title = "Distribution of image similarity"
) +
theme_minimal()
Main similarity plots
CLIP analysis: current similarity effects are dubious with lots of dataset level variance of course that will have to be accounted for in a mixed effects model; Garrison Bergelson dataset has individualized trials so difficult to make use of as well.
# N = the number of participants in a single trial here
adams_marchman_data_summarized <- summarize_similarity_data_collapsed(usable_trials_summarized_with_sims, extra_fields = c("dataset_name", "vanilla_trial")) |> filter(N > 10 & dataset_name == "adams_marchman_2018" & vanilla_trial==1)
am_plots <- generate_multimodal_plots(adams_marchman_data_summarized, "CLIP", title="Adams & Marchman, 2018")
am_plots
weaver_zettersten_data_summarized <- summarize_similarity_data_collapsed(usable_trials_summarized_with_sims, extra_fields = c("dataset_name", "vanilla_trial")) |> filter(N > 10 & dataset_name == "weaver_zettersten_2024" & vanilla_trial==1)
wz_plots <- generate_multimodal_plots(weaver_zettersten_data_summarized, "CLIP", title="Weaver et al., 2024")
wz_plots
clip_plots <- generate_multimodal_plots(clip_data_summarized |> filter(vanilla_trial==1), "CLIP", title="all vanilla trials")
clip_plots
generate_multimodal_plots(vanilla |> filter(N >= 5), "CLIP", title = "all vanilla trials (N >= 5)", size_by_n = TRUE, base_size=16)
more multimodal plot visualizations
generate_multimodal_plots(vanilla, "CLIP", title = "all vanilla trials", size_by_n = TRUE, base_size=16)
generate_multimodal_plots(vanilla, "CLIP", title = "by dataset", group_var = "dataset_name", se_alpha=0.2, point_alpha=0.1, show_legend = TRUE, line_alpha=0.4, show_cor=FALSE, overall_line=TRUE, overall_alpha=0.7, base_size=16)
generate_multimodal_plots(clip_data_summarized, "CLIP", title = "different trial types", group_var = "vanilla_trial", point_alpha=0.1, show_se=TRUE, base_size=16)
generate_multimodal_plots(clip_data_summarized_age_bucket |> filter(vanilla_trial == 1 & !is.na(age_bucket) & N>=5), "CLIP", title = "different age buckets (N>=5)", group_var = "age_bucket", show_se = TRUE, se_alpha = 0.2, point_alpha=0.1, line_alpha = 0.8, line_width=1.1, base_size=16)
clip_data_summarized_age_bucket_bigger <- summarize_similarity_data_collapsed(usable_trials_summarized_with_sims, extra_fields = c("dataset_name", "vanilla_trial", "aoa", "MeanSaliencyDiff", "age_bucket_bigger")) |> mutate(
sim_logit = qlogis(pmin(pmax(image_similarity, 1e-6), 1 - 1e-6)))
generate_multimodal_plots(clip_data_summarized_age_bucket_bigger |> filter(vanilla_trial == 1 & !is.na(age_bucket_bigger) & N>=5), "CLIP", title = "different larger age buckets (N>=5)", group_var = "age_bucket_bigger", show_se = TRUE, se_alpha = 0.2, point_alpha=0.1, line_alpha = 0.8, line_width=1.1, base_size=16)
By age
plot_data_age <- clip_data_summarized_age_bucket |>
pivot_longer(all_of(sims), names_to = "sim_type", values_to = "similarity") |>
mutate(sim_type = factor(sim_type, levels = sims))
ggplot(
plot_data_age |> filter(!is.na(age_bucket)),
aes(
similarity,
mean_value,
colour = age_bucket,
fill = age_bucket
)
) +
geom_hline(yintercept = 0.5, linetype = "dashed") +
geom_point(alpha = 0.1, size = 2) +
stat_smooth(
method = "lm",
se = TRUE,
aes(fill = age_bucket),
alpha = 0.15 # ribbon alpha
) +
stat_smooth(
method = "lm",
se = FALSE,
linewidth = 1.2
) +
facet_wrap(
~ sim_type,
nrow = 2,
scales = "free_x",
labeller = labeller(sim_type = facet_labs),
strip.position = "top"
) +
scale_color_discrete_qualitative(palette = "Set 3") +
labs(
x = "Target–distractor similarity",
y = "Proportion target looking in critical window",
colour = "Age bucket"
) +
guides(fill = "none") +
theme_classic(base_size = 16) +
theme(
legend.position = "bottom",
strip.text = element_text(size = 20),
panel.spacing = unit(1, "lines"),
plot.margin = margin(t = 5.5, r = 30, b = 5.5, l = 5.5, unit = "pt")
)
comparing all IVs
library(GGally)
confusability_vars <- clip_data_summarized |>
filter(vanilla_trial==1) |>
transmute(
mean_value=scale(mean_value)[, 1],
image_similarity=scale(image_similarity)[, 1],
text_similarity=scale(text_similarity)[, 1],
multimodal_similarity=scale(multimodal_similarity)[, 1],
ooo_similarity=scale(ooo_similarity)[, 1],
aoa=scale(aoa)[, 1],
MeanSaliencyDiff=scale(MeanSaliencyDiff)[, 1]
)
p <- ggpairs(
confusability_vars,
upper = list(continuous = wrap("cor", method = "spearman",
use = "pairwise.complete.obs", size = 3)),
lower = list(continuous = wrap("smooth", method = "lm",
alpha = 0.4, size = 0.8)),
diag = list(continuous = wrap("densityDiag", alpha = 0.5))
) +
theme_bw()
p
so many interesting colinear effects here with AoA and mean saliency..
# Load required libraries
library(ggcorrplot)
library(Hmisc) # Required for computing p-values
# 1. Prepare and rename your data
confusability_vars <- clip_data_summarized |>
filter(vanilla_trial == 1) |>
transmute(
# mean_value = scale(mean_value)[, 1],
image_similarity = scale(image_similarity)[, 1],
text_similarity = scale(text_similarity)[, 1],
multimodal_similarity = scale(multimodal_similarity)[, 1],
ooo_similarity = scale(ooo_similarity)[, 1],
aoa = scale(aoa)[, 1],
MeanSaliencyDiff = scale(MeanSaliencyDiff)[, 1]
) |>
# RENAME YOUR PREDICTORS HERE:
# Format: "New Name" = old_name
rename(
# "Critical window looking" = mean_value,
"Image similarity" = image_similarity,
"Text similarity" = text_similarity,
"Multimodal similarity" = multimodal_similarity,
"Adult behavioral similarity" = ooo_similarity,
"Estimated target age of acquisition" = aoa,
"Saliency difference" = MeanSaliencyDiff
)
cor_data <- rcorr(as.matrix(confusability_vars), type = "spearman")
# 3. Flatten matrices into a long dataframe for ggplot
r_df <- as.data.frame(cor_data$r) %>%
rownames_to_column(var = "Var1") %>%
pivot_longer(-Var1, names_to = "Var2", values_to = "r")
p_df <- as.data.frame(cor_data$P) %>%
rownames_to_column(var = "Var1") %>%
pivot_longer(-Var1, names_to = "Var2", values_to = "p")
# Combine r and p data, and create the significance star labels
plot_data <- left_join(r_df, p_df, by = c("Var1", "Var2")) %>%
# Fill diagonal p-values (which are NA) with 1 so they don't trigger stars
mutate(p = ifelse(Var1 == Var2, 1, p)) %>%
mutate(
stars = case_when(
p < 0.001 ~ "***",
p < 0.01 ~ "**",
p < 0.05 ~ "*",
TRUE ~ ""
),
# Combine the correlation number and stars into one text label
label_text = paste0(round(r, 2), stars)
)
# 4. Filter to get only the LOWER half of the square matrix
# We convert variables to factors to respect the order of our dataframe
vars_order <- colnames(confusability_vars)
plot_data <- plot_data %>%
mutate(
Var1 = factor(Var1, levels = vars_order),
Var2 = factor(Var2, levels = vars_order)
) %>%
filter(as.integer(Var1) > as.integer(Var2)) # Removes diagonal and upper triangle
# 5. Build the Heatmap using ggplot2
ggplot(plot_data, aes(x = Var2, y = Var1, fill = r)) +
geom_tile(color = "white") +
# Add the correlation values and significance stars combined
geom_text(aes(label = label_text), size = 3.5, color = "black") +
# Define the diverging color palette (Negative, Neutral, Positive)
scale_fill_gradient2(
low = "#6D9EC1", mid = "white", high = "#E46726",
midpoint = 0, limit = c(-1, 1), space = "Lab",
name = "Spearman\ncorrelation"
) +
theme_bw() +
labs(title = "Correlation of looking behavior predictors", x = NULL, y = NULL) +
theme(
axis.text.x = element_text(angle = 45, vjust = 1, hjust = 1),
panel.grid.major = element_blank(),
panel.border = element_blank(),
axis.ticks = element_blank()
)
ggsave(here("figures/correlation.png"))Analysis across age
library(dplyr)
calculate_correlations <- function(data, x_var, y_var, group_var = c("rounded_age"), conf_level = 0.95) {
data |>
dplyr::group_by(across(all_of(group_var))) |>
dplyr::summarize(
{
cor_test <- cor.test(.data[[x_var]], .data[[y_var]], method = "pearson", conf.level = conf_level)
tibble(
pearson_cor = cor_test$estimate,
p_value = cor_test$p.value,
ci_lower = cor_test$conf.int[1],
ci_upper = cor_test$conf.int[2]
)
},
.groups = "drop"
)
}
sim_age_plot <- function(data) {
ggplot(data, aes(x = rounded_age, y = pearson_cor, color = similarity_type)) +
geom_hline(yintercept = 0, linetype = "dashed") +
geom_point(size = 3, position = position_dodge(width=0.5)) + # Apply jitter to points only
geom_errorbar(aes(ymin = ci_lower, ymax = ci_upper),
width = 0.3, alpha = 0.2,
position=position_dodge(width=0.5)) + # No jitter on error bars
geom_smooth(span = 2, alpha=0.1, se=FALSE) +
labs(title = paste("Similarity correlations across age"),
x = "Age",
y = "Coefficient of similarity") +
theme_minimal() +
guides(shape = "none") +
scale_x_continuous(breaks=seq(5,70,5)) +
scale_color_brewer(palette = "Set1", name = "Similarity type")
}
# can't figure out what values to filter to here?
clip_data_age_summarized <- summarize_similarity_data_collapsed(age_based_trials |> filter(vanilla_trial == 1), extra_fields = c("rounded_age", "dataset_name")) |> filter(N >= 5) |> group_by(rounded_age) |>
filter(n() >= 5) |>
ungroup()
clip_age_image_cors <- calculate_correlations(clip_data_age_summarized, "image_similarity", "mean_value") |> mutate(similarity_type = "image")
clip_age_text_cors <- calculate_correlations(clip_data_age_summarized, "text_similarity", "mean_value") |> mutate(similarity_type = "text")
clip_age_multimodal_cors <- calculate_correlations(clip_data_age_summarized, "multimodal_similarity", "mean_value") |> mutate(similarity_type = "multimodal")
clip_age_ooo_cors <- calculate_correlations(clip_data_age_summarized, "ooo_similarity", "mean_value") |> mutate(similarity_type = "ooo")
clip_age_cors <- bind_rows(clip_age_image_cors, clip_age_text_cors, clip_age_multimodal_cors, clip_age_ooo_cors)
sim_age_plot(clip_age_cors)
ggplot(clip_age_cors, aes(x = rounded_age, y = pearson_cor)) +
geom_point(aes(color = p_value < 0.05), size = 3) +
geom_smooth(span = 2) +
labs(title = "Image similarity correlation across age",
x = "Age in months",
y = "Pearson Correlation") +
scale_color_manual(values = c("TRUE" = "black", "FALSE" = "gray")) + # Set color for significance
theme_minimal() +
theme(legend.position = "none")
time
library(tidyr)
crit <- qnorm(0.975)
# one row per trial type, median-split within each similarity dimension
sims_long <- usable_trials_summarized_with_sims |> filter(vanilla_trial == 1) |>
select(trial_type_id, ends_with("_similarity")) |>
distinct() |>
pivot_longer(ends_with("_similarity"),
names_to = "sim_type", values_to = "similarity") |>
filter(!is.na(similarity)) |>
group_by(sim_type) |>
mutate(sim_split = if_else(similarity >= median(similarity, na.rm = TRUE),
"high", "low")) |>
ungroup() |>
mutate(sim_type = sub("_similarity$", "", sim_type))
aoi_base <- downloaded_aois |>
filter(!is.na(accuracy), trial_type_id %in% sims_long$trial_type_id)
# fan out: each timepoint contributes to its high/low bucket in each dimension
aoi_split <- aoi_base |>
left_join(sims_long, by = "trial_type_id", relationship = "many-to-many")
# subject means per condition x time, then across-subject mean + 95% CI
across_split <- aoi_split |>
group_by(sim_type, sim_split, administration_id, t_norm) |>
summarise(mean_accuracy = mean(accuracy, na.rm = TRUE), .groups = "drop") |>
group_by(sim_type, sim_split, t_norm) |>
summarise(accuracy = mean(mean_accuracy, na.rm = TRUE),
ci = crit * sd(mean_accuracy, na.rm = TRUE) / sqrt(n()),
.groups = "drop") |>
mutate(condition = paste(sim_type, sim_split))
summarize_across_subj_aois <- readRDS(here("intermediates/summarize_across_subj_aois.Rds"))
# the existing overall line as the 9th condition
plot_df <- bind_rows(
across_split |> select(t_norm, accuracy, ci, condition),
summarize_across_subj_aois |> mutate(condition = "Overall") |>
select(t_norm, accuracy, ci, condition)
) |>
mutate(condition = forcats::fct_relevel(condition, "Overall"))
looking_times <- ggplot(plot_df, aes(t_norm, accuracy,
colour = condition, fill = condition)) +
xlim(-2000, 4000) +
geom_errorbar(aes(ymin = accuracy - ci, ymax = accuracy + ci),
width = 0, alpha = 0.15) +
geom_smooth(method = "gam", se = FALSE, linewidth = 1) +
geom_vline(xintercept = 0, size = 1.5) +
geom_hline(yintercept = 0.5, size = 1.2, linetype = "dashed") +
geom_vline(xintercept = 300, linetype = "dotted") +
ylim(0, 1) +
xlab("Time (normalized to target word onset) in ms") +
ylab("Proportion Target Looking") +
labs(colour = "Condition", fill = "Condition")
looking_times
stats
pre-registered models
library(lmerTest)
library(glmmTMB)
library(MuMIn)
library(broom.mixed)
model_data <- usable_trials_summarized_with_sims |> filter(vanilla_trial == 1)
model_data_scaled <- model_data |>
mutate(
# Scale your response variable but keep the original for plotting later
scaled_response = as.vector(scale(mean_target_looking_critical_window)),
scaled_age = as.vector(scale(age)),
scaled_aoa = as.vector(scale(aoa)),
scaled_saliency_diff = as.vector(scale(MeanSaliencyDiff))
) |>
# Dynamic scaling for your various similarity columns
mutate(across(all_of(sims), ~as.vector(scale(.x)), .names = "scaled_{.col}"))
fit_main <- function(sim,
data = model_data,
response = "mean_target_looking_critical_window",
added_structure = NULL, pruned_model=FALSE) {
terms <- c(
sprintf("scale(%s)*scale(age)", sim),
"scale(aoa)",
"scale(MeanSaliencyDiff)",
"(1 | original_target_label:img_key)",
"(1 | dataset_id)"
)
if (pruned_model) {
terms <- c(terms, "(1 | subject_id)")
} else {
terms <- c(terms, sprintf("(1 + scale(%s) | subject_id)", sim))
}
if (!is.null(added_structure)) {
terms <- c(terms, added_structure)
}
f <- reformulate(
terms,
response = sprintf("scale(%s)", response)
)
lmer(f, data = data)
}
fit_main_scaled <- function(sim, data = model_data_scaled, pruned_model = FALSE) {
# Use the pre-scaled variable names
terms <- c(
sprintf("scaled_%s * scaled_age", sim),
"scaled_aoa",
"scaled_saliency_diff",
"(1 | original_target_label:img_key)",
"(1 | dataset_id)"
)
if (pruned_model) {
terms <- c(terms, "(1 | subject_id)")
} else {
terms <- c(terms, sprintf("(1 + scaled_%s | subject_id)", sim))
}
# Response is already pre-scaled
f <- reformulate(terms, response = "scaled_response")
lmer(f, data = data)
}
mods <- lapply(sims, fit_main); names(mods) <- sims
mods_scaled <- lapply(sims, fit_main_scaled); names(mods_scaled) <- sims
lapply(mods, function(m) m@optinfo$conv$lme4$messages)$image_similarity
NULL
$text_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$multimodal_similarity
NULL
$ooo_similarity
NULLlapply(mods, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58356.9
Scaled residuals:
Min 1Q Median 3Q Max
-3.3201 -0.6271 0.1156 0.7219 2.4420
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.044743 0.21152
scale(image_similarity) 0.002014 0.04488 -0.18
original_target_label:img_key (Intercept) 0.051918 0.22786
dataset_id (Intercept) 0.022916 0.15138
Residual 0.850784 0.92238
Number of obs: 21312, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -5.716e-03 3.588e-02 2.214e+01 -0.159
scale(image_similarity) -3.231e-02 1.741e-02 4.442e+02 -1.856
scale(age) 2.106e-01 1.513e-02 8.437e+02 13.920
scale(aoa) 1.767e-03 1.189e-02 7.046e+02 0.149
scale(MeanSaliencyDiff) 6.220e-03 1.202e-02 4.642e+02 0.517
scale(image_similarity):scale(age) -6.794e-03 1.096e-02 2.275e+03 -0.620
Pr(>|t|)
(Intercept) 0.8749
scale(image_similarity) 0.0641 .
scale(age) <2e-16 ***
scale(aoa) 0.8819
scale(MeanSaliencyDiff) 0.6052
scale(image_similarity):scale(age) 0.5354
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mg_sml) -0.058
scale(age) -0.099 -0.044
scale(aoa) -0.077 -0.079 -0.045
scl(MnSlnD) 0.000 0.083 0.023 0.069
scl(mg_):() -0.033 0.064 -0.292 0.027 -0.051
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58359.1
Scaled residuals:
Min 1Q Median 3Q Max
-3.3210 -0.6265 0.1140 0.7246 2.4044
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.0444983 0.2109
scale(text_similarity) 0.0004709 0.0217 -1.00
original_target_label:img_key (Intercept) 0.0526953 0.2296
dataset_id (Intercept) 0.0233636 0.1529
Residual 0.8511638 0.9226
Number of obs: 21312, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -8.421e-03 3.666e-02 2.253e+01 -0.230
scale(text_similarity) 9.436e-04 1.726e-02 5.228e+02 0.055
scale(age) 2.095e-01 1.506e-02 8.929e+02 13.915
scale(aoa) 1.114e-03 1.217e-02 6.855e+02 0.092
scale(MeanSaliencyDiff) 7.590e-03 1.202e-02 4.681e+02 0.632
scale(text_similarity):scale(age) 5.362e-03 1.092e-02 3.014e+03 0.491
Pr(>|t|)
(Intercept) 0.820
scale(text_similarity) 0.956
scale(age) <2e-16 ***
scale(aoa) 0.927
scale(MeanSaliencyDiff) 0.528
scale(text_similarity):scale(age) 0.624
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(txt_sm) 0.148
scale(age) -0.088 0.015
scale(aoa) -0.042 0.188 -0.017
scl(MnSlnD) 0.010 0.032 0.018 0.083
scl(tx_):() 0.058 -0.110 0.251 0.051 0.023
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58357.3
Scaled residuals:
Min 1Q Median 3Q Max
-3.3228 -0.6271 0.1141 0.7202 2.4608
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.044987 0.21210
scale(multimodal_similarity) 0.003344 0.05782
original_target_label:img_key (Intercept) 0.051844 0.22769
dataset_id (Intercept) 0.022924 0.15141
Residual 0.848875 0.92134
Corr
-0.11
Number of obs: 21312, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -1.128e-02 3.582e-02 2.231e+01
scale(multimodal_similarity) -1.815e-02 1.346e-02 4.284e+02
scale(age) 2.071e-01 1.458e-02 7.226e+02
scale(aoa) -2.018e-03 1.202e-02 7.084e+02
scale(MeanSaliencyDiff) 7.140e-03 1.196e-02 4.698e+02
scale(multimodal_similarity):scale(age) 1.628e-03 9.464e-03 2.227e+03
t value Pr(>|t|)
(Intercept) -0.315 0.756
scale(multimodal_similarity) -1.348 0.178
scale(age) 14.207 <2e-16 ***
scale(aoa) -0.168 0.867
scale(MeanSaliencyDiff) 0.597 0.551
scale(multimodal_similarity):scale(age) 0.172 0.863
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mltmd_) 0.022
scale(age) -0.113 0.005
scale(aoa) -0.075 0.150 -0.037
scl(MnSlnD) 0.003 0.024 0.008 0.080
scl(ml_):() 0.015 -0.036 0.112 0.016 -0.027
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 55456.5
Scaled residuals:
Min 1Q Median 3Q Max
-3.3359 -0.6230 0.1168 0.7207 2.4523
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.046212 0.21497
scale(ooo_similarity) 0.005711 0.07557 -0.20
original_target_label:img_key (Intercept) 0.048499 0.22022
dataset_id (Intercept) 0.020838 0.14435
Residual 0.840619 0.91685
Number of obs: 20316, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -1.404e-02 3.498e-02 2.194e+01 -0.401
scale(ooo_similarity) -2.645e-02 1.704e-02 3.537e+02 -1.552
scale(age) 2.139e-01 1.492e-02 5.606e+02 14.334
scale(aoa) -3.091e-03 1.229e-02 6.394e+02 -0.252
scale(MeanSaliencyDiff) 7.709e-03 1.223e-02 4.242e+02 0.630
scale(ooo_similarity):scale(age) 1.665e-02 9.836e-03 1.769e+03 1.693
Pr(>|t|)
(Intercept) 0.6921
scale(ooo_similarity) 0.1215
scale(age) <2e-16 ***
scale(aoa) 0.8015
scale(MeanSaliencyDiff) 0.5289
scale(ooo_similarity):scale(age) 0.0907 .
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(_smlrt) 0.103
scale(age) -0.125 -0.028
scale(aoa) -0.101 -0.079 -0.040
scl(MnSlnD) 0.015 0.064 0.006 0.070
scl(_sm):() -0.038 -0.127 0.196 0.026 -0.030r.squaredGLMM(mods$image_similarity) R2m R2c
[1,] 0.04098928 0.1610329Trying to see why text and multimodal similarity models were singular.
vars <- c("image_similarity", "text_similarity", "multimodal_similarity",
"ooo_similarity", "mean_target_looking_critical_window",
"age", "aoa", "MeanSaliencyDiff")
colSums(is.na(model_data[vars])) image_similarity text_similarity
0 0
multimodal_similarity ooo_similarity
0 1118
mean_target_looking_critical_window age
0 131
aoa MeanSaliencyDiff
1260 0 # total rows, and rows complete across all vars
nrow(model_data)[1] 22703sum(complete.cases(model_data[vars]))[1] 20316# rows complete without requiring ooo
sum(complete.cases(model_data[setdiff(vars, "ooo_similarity")]))[1] 21312library(ggeffects)
mu <- mean(model_data$mean_target_looking_critical_window, na.rm = TRUE)
sig <- sd(model_data$mean_target_looking_critical_window, na.rm = TRUE)
all_predictions <- lapply(sims, function(sim) {
term_name <- sim
# Generate prediction
pred <- ggpredict(mods[[sim]], terms = term_name)
# Label it so we know which facet it belongs to
pred$sim_type <- paste(gsub("_", " ", sim))
return(pred)
})
combined_main_preds <- do.call(rbind, all_predictions)
combined_main_preds$predicted_raw <- combined_main_preds$predicted * sig + mu
combined_main_preds$conf.low_raw <- combined_main_preds$conf.low * sig + mu
combined_main_preds$conf.high_raw <- combined_main_preds$conf.high * sig + mu
ggplot(
combined_main_preds,
aes(x = x, y = predicted_raw)
) +
# Confidence intervals ribbon
geom_ribbon(
aes(ymin = conf.low_raw, ymax = conf.high_raw),
alpha = 0.15,
fill = "steelblue"
) +
# Main effect prediction line
geom_line(
linewidth = 1.2,
color = "steelblue"
) +
# 50% chance line (useful for target looking paradigms)
geom_hline(
yintercept = 0.5,
linetype = "dashed",
alpha = 0.5
) +
# Separate into 4 panels side-by-side / wrapped
facet_wrap(~ sim_type, scales = "free_x") +
labs(
title = "Main similarity predictors",
subtitle = "Critical Window Predictions",
x = "Target-distractor similarity",
y = "Predicted Proportion Target Looking"
) +
theme_classic() +
theme(
strip.background = element_blank(),
strip.text = element_text(face = "bold", size = 11)
)
getting rid of singular effects
getting rid of singular effects
pruned_text_model <- lmer(scale(mean_target_looking_critical_window) ~ scale(text_similarity)*scale(age)
+ scale(aoa) + scale(MeanSaliencyDiff)
+ (1 | subject_id)
+ (1 | original_target_label:img_key)
+ (1 | dataset_id), data = model_data)
summary(pruned_text_model)Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: scale(mean_target_looking_critical_window) ~ scale(text_similarity) *
scale(age) + scale(aoa) + scale(MeanSaliencyDiff) + (1 |
subject_id) + (1 | original_target_label:img_key) + (1 | dataset_id)
Data: model_data
REML criterion at convergence: 58365
Scaled residuals:
Min 1Q Median 3Q Max
-3.3184 -0.6273 0.1148 0.7229 2.4557
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04517 0.2125
original_target_label:img_key (Intercept) 0.05270 0.2296
dataset_id (Intercept) 0.02383 0.1544
Residual 0.85203 0.9231
Number of obs: 21312, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -7.990e-03 3.693e-02 2.283e+01 -0.216
scale(text_similarity) 1.656e-03 1.715e-02 5.306e+02 0.097
scale(age) 2.089e-01 1.509e-02 9.302e+02 13.842
scale(aoa) 6.770e-04 1.214e-02 6.870e+02 0.056
scale(MeanSaliencyDiff) 7.405e-03 1.201e-02 4.684e+02 0.616
scale(text_similarity):scale(age) 4.792e-03 1.088e-02 3.133e+03 0.441
Pr(>|t|)
(Intercept) 0.831
scale(text_similarity) 0.923
scale(age) <2e-16 ***
scale(aoa) 0.956
scale(MeanSaliencyDiff) 0.538
scale(text_similarity):scale(age) 0.660
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(txt_sm) 0.158
scale(age) -0.086 0.011
scale(aoa) -0.042 0.189 -0.016
scl(MnSlnD) 0.010 0.031 0.018 0.084
scl(tx_):() 0.053 -0.105 0.278 0.051 0.024model comparison
Effects persist in the pruned models. Now trying to compare to see which one has the best fit.
dat_cc <- model_data |>
tidyr::drop_na(image_similarity, text_similarity, multimodal_similarity, ooo_similarity,
mean_target_looking_critical_window,
age, aoa, MeanSaliencyDiff)
fit_common <- function(sim, data=dat_cc) {
f <- reformulate(
c(sprintf("scale(%s)*scale(age)", sim),
"scale(aoa)", "scale(MeanSaliencyDiff)",
"(1 | subject_id)",
"(1 | original_target_label:img_key)",
"(1 | dataset_id)"),
response = "scale(mean_target_looking_critical_window)")
lmer(f, data = dat_cc, REML = FALSE)
}
mods_common <- lapply(sims, fit_common); names(mods_common) <- sims
sel <- model.sel(mods_common$image_similarity, mods_common$text_similarity,
mods_common$multimodal_similarity, mods_common$ooo_similarity)
selModel selection table
(Int) scl(age) scl(aoa) scl(img_sml)
mods_common$ooo_similarity -0.02318 0.2185 -0.003486
mods_common$image_similarity -0.01094 0.2172 -0.003830 -0.03736
mods_common$multimodal_similarity -0.01786 0.2139 -0.007671
mods_common$text_similarity -0.01382 0.2153 -0.004190
scl(MSD) scl(age):scl(img_sml) scl(txt_sml)
mods_common$ooo_similarity 0.007068
mods_common$image_similarity 0.007010 -0.006103
mods_common$multimodal_similarity 0.008287
mods_common$text_similarity 0.008957 0.005492
scl(age):scl(txt_sml) scl(mlt_sml)
mods_common$ooo_similarity
mods_common$image_similarity
mods_common$multimodal_similarity -0.01998
mods_common$text_similarity 0.004236
scl(age):scl(mlt_sml) scl(ooo_sml)
mods_common$ooo_similarity -0.02563
mods_common$image_similarity
mods_common$multimodal_similarity 0.003695
mods_common$text_similarity
scl(age):scl(ooo_sml) df logLik AICc
mods_common$ooo_similarity 0.01669 10 -27807.11 55634.2
mods_common$image_similarity 10 -27807.25 55634.5
mods_common$multimodal_similarity 10 -27808.33 55636.7
mods_common$text_similarity 10 -27809.32 55638.7
delta weight
mods_common$ooo_similarity 0.00 0.441
mods_common$image_similarity 0.29 0.382
mods_common$multimodal_similarity 2.45 0.130
mods_common$text_similarity 4.43 0.048
Models ranked by AICc(x)
Random terms (all models):
1 | subject_id, 1 | original_target_label:img_key, 1 | dataset_idsummary(mods_common$image_similarity)Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite's
method [lmerModLmerTest]
Formula: f
Data: dat_cc
AIC BIC logLik -2*log(L) df.resid
55634.5 55713.7 -27807.3 55614.5 20306
Scaled residuals:
Min 1Q Median 3Q Max
-3.3407 -0.6238 0.1154 0.7224 2.4765
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04875 0.2208
original_target_label:img_key (Intercept) 0.04813 0.2194
dataset_id (Intercept) 0.01917 0.1385
Residual 0.85235 0.9232
Number of obs: 20316, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -1.094e-02 3.380e-02 2.310e+01 -0.324
scale(image_similarity) -3.736e-02 1.795e-02 4.085e+02 -2.082
scale(age) 2.172e-01 1.547e-02 7.556e+02 14.035
scale(aoa) -3.830e-03 1.217e-02 6.345e+02 -0.315
scale(MeanSaliencyDiff) 7.010e-03 1.234e-02 4.252e+02 0.568
scale(image_similarity):scale(age) -6.103e-03 1.164e-02 2.163e+03 -0.525
Pr(>|t|)
(Intercept) 0.749
scale(image_similarity) 0.038 *
scale(age) <2e-16 ***
scale(aoa) 0.753
scale(MeanSaliencyDiff) 0.570
scale(image_similarity):scale(age) 0.600
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mg_sml) -0.080
scale(age) -0.091 -0.060
scale(aoa) -0.094 -0.064 -0.054
scl(MnSlnD) 0.006 0.082 0.026 0.067
scl(mg_):() -0.047 0.088 -0.312 0.033 -0.051Image similarity model comes out on top.
including non-vanilla trials
model_data_all <- usable_trials_summarized_with_sims |>
mutate(vanilla_trial = ifelse(vanilla_trial==1, 0.5, -0.5))
mods_all <- lapply(sims, fit_main, model_data_all, added_structure="scale(vanilla_trial)"); names(mods_all) <- sims
lapply(mods_all, function(m) m@optinfo$conv$lme4$messages)$image_similarity
NULL
$text_similarity
NULL
$multimodal_similarity
NULL
$ooo_similarity
NULLlapply(mods_all, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 73188.1
Scaled residuals:
Min 1Q Median 3Q Max
-3.4171 -0.6169 0.1141 0.7050 2.5167
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.043822 0.20934
scale(image_similarity) 0.002493 0.04993 -0.15
original_target_label:img_key (Intercept) 0.069699 0.26401
dataset_id (Intercept) 0.030970 0.17598
Residual 0.811356 0.90075
Number of obs: 27109, groups:
subject_id, 1579; original_target_label:img_key, 1097; dataset_id, 26
Fixed effects:
Estimate Std. Error df t value
(Intercept) 5.764e-02 3.810e-02 2.380e+01 1.513
scale(image_similarity) -4.204e-02 1.713e-02 7.114e+02 -2.454
scale(age) 2.203e-01 1.539e-02 1.332e+03 14.311
scale(aoa) 1.295e-03 1.139e-02 8.538e+02 0.114
scale(MeanSaliencyDiff) 2.110e-02 1.123e-02 7.200e+02 1.879
scale(vanilla_trial) 3.172e-03 1.433e-02 1.427e+03 0.221
scale(image_similarity):scale(age) -2.688e-02 1.055e-02 2.580e+03 -2.549
Pr(>|t|)
(Intercept) 0.1435
scale(image_similarity) 0.0144 *
scale(age) <2e-16 ***
scale(aoa) 0.9095
scale(MeanSaliencyDiff) 0.0606 .
scale(vanilla_trial) 0.8248
scale(image_similarity):scale(age) 0.0109 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(m_) scl(g) scal() s(MSD) scl(v_)
scl(mg_sml) 0.016
scale(age) 0.024 -0.091
scale(aoa) -0.054 -0.044 -0.025
scl(MnSlnD) 0.000 0.063 0.035 0.000
scl(vnll_t) 0.030 0.112 0.022 -0.050 0.005
scl(mg_):() -0.047 0.113 -0.328 0.034 -0.056 0.032
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 71722.2
Scaled residuals:
Min 1Q Median 3Q Max
-3.3978 -0.6188 0.1182 0.7065 2.4744
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.0443166 0.21051
scale(text_similarity) 0.0004793 0.02189 -0.85
original_target_label:img_key (Intercept) 0.0627652 0.25053
dataset_id (Intercept) 0.0321686 0.17936
Residual 0.8177373 0.90429
Number of obs: 26522, groups:
subject_id, 1579; original_target_label:img_key, 1090; dataset_id, 26
Fixed effects:
Estimate Std. Error df t value
(Intercept) 6.189e-02 3.868e-02 2.350e+01 1.600
scale(text_similarity) 2.860e-03 1.715e-02 6.727e+02 0.167
scale(age) 2.069e-01 1.510e-02 1.347e+03 13.707
scale(aoa) 7.681e-04 1.135e-02 7.922e+02 0.068
scale(MeanSaliencyDiff) 2.164e-02 1.085e-02 6.743e+02 1.993
scale(vanilla_trial) -7.525e-03 1.445e-02 1.373e+03 -0.521
scale(text_similarity):scale(age) -6.710e-03 1.200e-02 3.204e+03 -0.559
Pr(>|t|)
(Intercept) 0.1230
scale(text_similarity) 0.8676
scale(age) <2e-16 ***
scale(aoa) 0.9461
scale(MeanSaliencyDiff) 0.0466 *
scale(vanilla_trial) 0.6026
scale(text_similarity):scale(age) 0.5761
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(t_) scl(g) scal() s(MSD) scl(v_)
scl(txt_sm) 0.077
scale(age) 0.013 -0.013
scale(aoa) -0.030 0.197 -0.016
scl(MnSlnD) -0.001 0.030 0.023 0.011
scl(vnll_t) 0.037 0.128 0.047 -0.010 0.005
scl(tx_):() 0.023 -0.123 0.070 0.018 0.013 0.115
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 73187
Scaled residuals:
Min 1Q Median 3Q Max
-3.4146 -0.6167 0.1124 0.7055 2.5370
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.043531 0.20864
scale(multimodal_similarity) 0.003853 0.06207
original_target_label:img_key (Intercept) 0.068580 0.26188
dataset_id (Intercept) 0.032965 0.18156
Residual 0.810517 0.90029
Corr
-0.11
Number of obs: 27109, groups:
subject_id, 1579; original_target_label:img_key, 1097; dataset_id, 26
Fixed effects:
Estimate Std. Error df
(Intercept) 4.357e-02 3.928e-02 2.492e+01
scale(multimodal_similarity) -4.470e-02 1.576e-02 7.191e+02
scale(age) 2.026e-01 1.458e-02 1.416e+03
scale(aoa) -3.613e-03 1.142e-02 8.592e+02
scale(MeanSaliencyDiff) 2.074e-02 1.112e-02 7.210e+02
scale(vanilla_trial) -7.697e-03 1.493e-02 1.639e+03
scale(multimodal_similarity):scale(age) -2.414e-02 1.076e-02 3.289e+03
t value Pr(>|t|)
(Intercept) 1.109 0.27799
scale(multimodal_similarity) -2.836 0.00469 **
scale(age) 13.897 < 2e-16 ***
scale(aoa) -0.316 0.75183
scale(MeanSaliencyDiff) 1.865 0.06266 .
scale(vanilla_trial) -0.515 0.60632
scale(multimodal_similarity):scale(age) -2.244 0.02488 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(m_) scl(g) scal() s(MSD) scl(v_)
scl(mltmd_) 0.111
scale(age) 0.006 -0.027
scale(aoa) -0.038 0.102 -0.014
scl(MnSlnD) -0.002 0.020 0.020 0.006
scl(vnll_t) 0.056 0.273 0.042 -0.007 0.003
scl(ml_):() -0.013 0.023 0.086 0.060 -0.014 0.140
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 62585.3
Scaled residuals:
Min 1Q Median 3Q Max
-3.3561 -0.6192 0.1124 0.7158 2.4673
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.045157 0.21250
scale(ooo_similarity) 0.005855 0.07652 -0.14
original_target_label:img_key (Intercept) 0.057803 0.24042
dataset_id (Intercept) 0.021789 0.14761
Residual 0.833943 0.91320
Number of obs: 22966, groups:
subject_id, 1544; original_target_label:img_key, 881; dataset_id, 26
Fixed effects:
Estimate Std. Error df t value
(Intercept) 4.483e-02 3.422e-02 2.191e+01 1.310
scale(ooo_similarity) -2.366e-02 1.704e-02 3.720e+02 -1.388
scale(age) 2.201e-01 1.575e-02 7.027e+02 13.976
scale(aoa) 1.931e-03 1.440e-02 6.728e+02 0.134
scale(MeanSaliencyDiff) 1.109e-02 1.243e-02 4.581e+02 0.892
scale(vanilla_trial) 1.976e-03 1.791e-02 1.306e+03 0.110
scale(ooo_similarity):scale(age) 2.539e-03 1.060e-02 2.032e+03 0.240
Pr(>|t|)
(Intercept) 0.204
scale(ooo_similarity) 0.166
scale(age) <2e-16 ***
scale(aoa) 0.893
scale(MeanSaliencyDiff) 0.373
scale(vanilla_trial) 0.912
scale(ooo_similarity):scale(age) 0.811
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_s) scl(g) scal() s(MSD) scl(v_)
scl(_smlrt) 0.109
scale(age) 0.010 -0.042
scale(aoa) -0.070 -0.069 -0.033
scl(MnSlnD) 0.019 0.048 0.014 0.058
scl(vnll_t) -0.106 0.087 0.016 -0.078 -0.029
scl(_sm):() -0.039 0.005 0.121 0.028 -0.019 0.061 multimodal similarity and image similarity ar the only significant ones here…pruning
pruned_multimodal_model_all <- lmer(scale(mean_target_looking_critical_window) ~ scale(multimodal_similarity)*scale(age)
+ scale(aoa) + scale(MeanSaliencyDiff)
+ (1 | subject_id)
+ (1|dataset_id), data = usable_trials_summarized_with_sims)
summary(pruned_multimodal_model_all)Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula:
scale(mean_target_looking_critical_window) ~ scale(multimodal_similarity) *
scale(age) + scale(aoa) + scale(MeanSaliencyDiff) + (1 |
subject_id) + (1 | dataset_id)
Data: usable_trials_summarized_with_sims
REML criterion at convergence: 74160.8
Scaled residuals:
Min 1Q Median 3Q Max
-3.0210 -0.6286 0.1237 0.7363 2.1261
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04349 0.2085
dataset_id (Intercept) 0.03377 0.1838
Residual 0.87136 0.9335
Number of obs: 27109, groups: subject_id, 1579; dataset_id, 26
Fixed effects:
Estimate Std. Error df
(Intercept) 3.696e-02 3.785e-02 2.418e+01
scale(multimodal_similarity) -4.924e-02 7.308e-03 2.633e+04
scale(age) 1.968e-01 1.444e-02 1.343e+03
scale(aoa) -4.212e-03 6.560e-03 2.451e+04
scale(MeanSaliencyDiff) 2.719e-02 5.874e-03 2.684e+04
scale(multimodal_similarity):scale(age) -2.884e-02 7.327e-03 2.602e+04
t value Pr(>|t|)
(Intercept) 0.976 0.339
scale(multimodal_similarity) -6.738 1.64e-11 ***
scale(age) 13.624 < 2e-16 ***
scale(aoa) -0.642 0.521
scale(MeanSaliencyDiff) 4.629 3.69e-06 ***
scale(multimodal_similarity):scale(age) -3.935 8.33e-05 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mltmd_) 0.046
scale(age) 0.000 -0.052
scale(aoa) -0.023 0.063 -0.006
scl(MnSlnD) 0.004 0.016 0.014 -0.041
scl(ml_):() -0.014 0.002 0.052 0.100 -0.071effect stays and with a pretty robust random effects structure and with an interaction with age. feels like this is a really interesting effect, especially given that the other models across vanilla and non-vanilla are not similarly predictive.
including baseline window as covariate
baseline_data <- model_data |> filter(min_time <= -500) # getting rid of 5 individual trials ha
mods_baseline_covariate <- lapply(sims, fit_main, data=baseline_data, added_structure="scale(mean_target_looking_baseline_window)", pruned_model=TRUE); names(mods_baseline_covariate) <- sims
lapply(mods_baseline_covariate, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58111.8
Scaled residuals:
Min 1Q Median 3Q Max
-3.2500 -0.6312 0.1191 0.7318 2.5275
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04618 0.2149
original_target_label:img_key (Intercept) 0.04152 0.2038
dataset_id (Intercept) 0.02398 0.1548
Residual 0.84431 0.9189
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -6.207e-03 3.611e-02 2.216e+01
scale(image_similarity) -3.260e-02 1.615e-02 4.290e+02
scale(age) 2.088e-01 1.506e-02 8.574e+02
scale(aoa) 1.208e-03 1.130e-02 6.911e+02
scale(MeanSaliencyDiff) 4.249e-03 1.127e-02 4.460e+02
scale(mean_target_looking_baseline_window) 1.033e-01 6.550e-03 2.085e+04
scale(image_similarity):scale(age) -7.285e-03 1.075e-02 2.624e+03
t value Pr(>|t|)
(Intercept) -0.172 0.8651
scale(image_similarity) -2.018 0.0442 *
scale(age) 13.863 <2e-16 ***
scale(aoa) 0.107 0.9149
scale(MeanSaliencyDiff) 0.377 0.7064
scale(mean_target_looking_baseline_window) 15.764 <2e-16 ***
scale(image_similarity):scale(age) -0.678 0.4980
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) s(____
scl(mg_sml) -0.049
scale(age) -0.102 -0.046
scale(aoa) -0.073 -0.078 -0.043
scl(MnSlnD) 0.001 0.084 0.022 0.068
scl(mn____) -0.002 -0.002 -0.007 0.000 -0.017
scl(mg_):() -0.033 0.078 -0.282 0.026 -0.053 -0.001
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58115.9
Scaled residuals:
Min 1Q Median 3Q Max
-3.2423 -0.6312 0.1179 0.7338 2.5262
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04611 0.2147
original_target_label:img_key (Intercept) 0.04238 0.2059
dataset_id (Intercept) 0.02494 0.1579
Residual 0.84424 0.9188
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -8.100e-03 3.715e-02 2.282e+01
scale(text_similarity) 3.128e-03 1.628e-02 5.257e+02
scale(age) 2.069e-01 1.506e-02 9.285e+02
scale(aoa) 4.049e-04 1.154e-02 6.700e+02
scale(MeanSaliencyDiff) 5.867e-03 1.129e-02 4.497e+02
scale(mean_target_looking_baseline_window) 1.031e-01 6.551e-03 2.085e+04
scale(text_similarity):scale(age) 4.337e-03 1.063e-02 2.813e+03
t value Pr(>|t|)
(Intercept) -0.218 0.829
scale(text_similarity) 0.192 0.848
scale(age) 13.734 <2e-16 ***
scale(aoa) 0.035 0.972
scale(MeanSaliencyDiff) 0.520 0.603
scale(mean_target_looking_baseline_window) 15.736 <2e-16 ***
scale(text_similarity):scale(age) 0.408 0.683
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) s(____
scl(txt_sm) 0.149
scale(age) -0.091 0.005
scale(aoa) -0.041 0.180 -0.015
scl(MnSlnD) 0.010 0.030 0.017 0.083
scl(mn____) -0.003 0.000 -0.009 0.000 -0.017
scl(tx_):() 0.049 -0.120 0.272 0.051 0.023 -0.005
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 58115.2
Scaled residuals:
Min 1Q Median 3Q Max
-3.2413 -0.6308 0.1185 0.7328 2.5241
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04613 0.2148
original_target_label:img_key (Intercept) 0.04197 0.2049
dataset_id (Intercept) 0.02408 0.1552
Residual 0.84431 0.9189
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -1.160e-02 3.614e-02 2.235e+01
scale(multimodal_similarity) -1.688e-02 1.240e-02 4.063e+02
scale(age) 2.050e-01 1.457e-02 7.379e+02
scale(aoa) -2.726e-03 1.142e-02 6.888e+02
scale(MeanSaliencyDiff) 5.362e-03 1.125e-02 4.513e+02
scale(mean_target_looking_baseline_window) 1.031e-01 6.551e-03 2.085e+04
scale(multimodal_similarity):scale(age) -3.357e-04 9.045e-03 2.767e+03
t value Pr(>|t|)
(Intercept) -0.321 0.751
scale(multimodal_similarity) -1.362 0.174
scale(age) 14.072 <2e-16 ***
scale(aoa) -0.239 0.811
scale(MeanSaliencyDiff) 0.476 0.634
scale(mean_target_looking_baseline_window) 15.738 <2e-16 ***
scale(multimodal_similarity):scale(age) -0.037 0.970
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) s(____
scl(mltmd_) 0.024
scale(age) -0.114 0.004
scale(aoa) -0.071 0.151 -0.034
scl(MnSlnD) 0.004 0.026 0.007 0.079
scl(mn____) -0.002 0.007 -0.009 0.001 -0.016
scl(ml_):() 0.012 -0.031 0.126 0.018 -0.029 -0.012
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 55255.3
Scaled residuals:
Min 1Q Median 3Q Max
-3.2374 -0.6265 0.1198 0.7296 2.5432
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04912 0.2216
original_target_label:img_key (Intercept) 0.03987 0.1997
dataset_id (Intercept) 0.02217 0.1489
Residual 0.83758 0.9152
Number of obs: 20311, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -1.482e-02 3.540e-02 2.225e+01
scale(ooo_similarity) -2.564e-02 1.593e-02 3.838e+02
scale(age) 2.125e-01 1.485e-02 6.062e+02
scale(aoa) -3.471e-03 1.172e-02 6.259e+02
scale(MeanSaliencyDiff) 5.299e-03 1.154e-02 4.080e+02
scale(mean_target_looking_baseline_window) 9.605e-02 6.656e-03 1.984e+04
scale(ooo_similarity):scale(age) 1.632e-02 9.471e-03 2.306e+03
t value Pr(>|t|)
(Intercept) -0.419 0.679
scale(ooo_similarity) -1.610 0.108
scale(age) 14.309 <2e-16 ***
scale(aoa) -0.296 0.767
scale(MeanSaliencyDiff) 0.459 0.646
scale(mean_target_looking_baseline_window) 14.430 <2e-16 ***
scale(ooo_similarity):scale(age) 1.723 0.085 .
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) s(____
scl(_smlrt) 0.106
scale(age) -0.126 -0.031
scale(aoa) -0.096 -0.079 -0.036
scl(MnSlnD) 0.015 0.070 0.005 0.068
scl(mn____) -0.002 0.001 -0.008 0.003 -0.016
scl(_sm):() -0.042 -0.128 0.222 0.028 -0.032 -0.003predicting baseline corrected looking
mods_baseline_corrected <- lapply(sims, fit_main, data=baseline_data, response="corrected_target_looking"); names(mods_baseline_corrected) <- sims
lapply(mods_baseline_corrected, function(m) m@optinfo$conv$lme4$messages)$image_similarity
NULL
$text_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$multimodal_similarity
NULL
$ooo_similarity
[1] "boundary (singular) fit: see help('isSingular')"lapply(mods_baseline_corrected, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59514.2
Scaled residuals:
Min 1Q Median 3Q Max
-3.11900 -0.63915 -0.02499 0.70133 2.46636
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.016818 0.12968
scale(image_similarity) 0.001158 0.03403 -0.33
original_target_label:img_key (Intercept) 0.007625 0.08732
dataset_id (Intercept) 0.012481 0.11172
Residual 0.932947 0.96589
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -1.116e-02 2.629e-02 2.178e+01 -0.425
scale(image_similarity) -3.007e-02 1.128e-02 2.490e+02 -2.665
scale(age) 1.326e-01 1.358e-02 4.144e+02 9.764
scale(aoa) -2.360e-03 8.766e-03 7.300e+02 -0.269
scale(MeanSaliencyDiff) -1.516e-02 8.168e-03 4.546e+02 -1.856
scale(image_similarity):scale(age) -4.733e-03 9.542e-03 1.087e+03 -0.496
Pr(>|t|)
(Intercept) 0.6753
scale(image_similarity) 0.0082 **
scale(age) <2e-16 ***
scale(aoa) 0.7878
scale(MeanSaliencyDiff) 0.0641 .
scale(image_similarity):scale(age) 0.6200
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mg_sml) -0.053
scale(age) -0.133 -0.057
scale(aoa) -0.072 -0.084 -0.053
scl(MnSlnD) 0.005 0.089 0.025 0.053
scl(mg_):() -0.037 0.118 -0.280 0.020 -0.063
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59522.9
Scaled residuals:
Min 1Q Median 3Q Max
-3.11902 -0.63720 -0.02521 0.70026 2.45410
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.017710 0.13308
scale(text_similarity) 0.000184 0.01356 -1.00
original_target_label:img_key (Intercept) 0.008458 0.09197
dataset_id (Intercept) 0.012059 0.10981
Residual 0.932676 0.96575
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -1.592e-02 2.632e-02 2.213e+01 -0.605
scale(text_similarity) -3.126e-03 1.201e-02 4.584e+02 -0.260
scale(age) 1.298e-01 1.355e-02 3.812e+02 9.581
scale(aoa) -4.342e-03 8.960e-03 7.155e+02 -0.485
scale(MeanSaliencyDiff) -1.348e-02 8.229e-03 4.581e+02 -1.638
scale(text_similarity):scale(age) 3.062e-04 9.219e-03 1.108e+03 0.033
Pr(>|t|)
(Intercept) 0.551
scale(text_similarity) 0.795
scale(age) <2e-16 ***
scale(aoa) 0.628
scale(MeanSaliencyDiff) 0.102
scale(text_similarity):scale(age) 0.974
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(txt_sm) 0.137
scale(age) -0.124 -0.004
scale(aoa) -0.048 0.126 -0.024
scl(MnSlnD) 0.011 0.021 0.015 0.068
scl(tx_):() 0.047 -0.215 0.257 0.076 0.015
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59521.4
Scaled residuals:
Min 1Q Median 3Q Max
-3.12403 -0.63665 -0.02594 0.70257 2.47284
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.0173206 0.13161
scale(multimodal_similarity) 0.0009835 0.03136
original_target_label:img_key (Intercept) 0.0084057 0.09168
dataset_id (Intercept) 0.0122593 0.11072
Residual 0.9323702 0.96559
Corr
0.33
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -0.015687 0.026116 21.784115
scale(multimodal_similarity) -0.005321 0.008845 315.218927
scale(age) 0.127943 0.013106 341.197657
scale(aoa) -0.005354 0.008904 715.338540
scale(MeanSaliencyDiff) -0.013001 0.008233 462.491157
scale(multimodal_similarity):scale(age) -0.010106 0.007859 978.700314
t value Pr(>|t|)
(Intercept) -0.601 0.554
scale(multimodal_similarity) -0.602 0.548
scale(age) 9.762 <2e-16 ***
scale(aoa) -0.601 0.548
scale(MeanSaliencyDiff) -1.579 0.115
scale(multimodal_similarity):scale(age) -1.286 0.199
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mltmd_) 0.029
scale(age) -0.148 0.001
scale(aoa) -0.073 0.131 -0.049
scl(MnSlnD) 0.006 0.032 0.004 0.067
scl(ml_):() 0.014 -0.055 0.117 0.018 -0.055
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 56921.8
Scaled residuals:
Min 1Q Median 3Q Max
-3.06679 -0.63558 -0.02463 0.70568 2.42243
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.0172670 0.13140
scale(ooo_similarity) 0.0002787 0.01670 -1.00
original_target_label:img_key (Intercept) 0.0072348 0.08506
dataset_id (Intercept) 0.0114660 0.10708
Residual 0.9422191 0.97068
Number of obs: 20311, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) -1.988e-02 2.575e-02 2.176e+01 -0.772
scale(ooo_similarity) -2.121e-02 1.154e-02 3.738e+02 -1.838
scale(age) 1.362e-01 1.333e-02 2.847e+02 10.215
scale(aoa) -2.259e-03 9.117e-03 6.766e+02 -0.248
scale(MeanSaliencyDiff) -1.446e-02 8.356e-03 4.128e+02 -1.730
scale(ooo_similarity):scale(age) 1.597e-02 8.090e-03 1.003e+03 1.974
Pr(>|t|)
(Intercept) 0.4484
scale(ooo_similarity) 0.0668 .
scale(age) <2e-16 ***
scale(aoa) 0.8044
scale(MeanSaliencyDiff) 0.0843 .
scale(ooo_similarity):scale(age) 0.0487 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(_smlrt) 0.100
scale(age) -0.158 -0.028
scale(aoa) -0.099 -0.112 -0.050
scl(MnSlnD) 0.021 0.078 0.004 0.050
scl(_sm):() -0.042 -0.196 0.183 0.038 -0.039
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')predicting window type
baseline_data_window_type <- baseline_data |> pivot_longer(cols=c(mean_target_looking_critical_window, mean_target_looking_baseline_window), names_to="window_type", values_to="target_looking") |>
mutate(window_type = str_replace(window_type, "mean_target_looking_", "")) |>
mutate(trial_window_c = case_when(
window_type=="critical_window" ~ 0.5,
window_type=="baseline_window" ~ -0.5))
mods_baseline_window_type <- lapply(sims, function(sim) {
fit_main(
sim = sim,
data = baseline_data_window_type,
response = "target_looking",
pruned_model = TRUE,
added_structure = sprintf("scale(trial_window_c) * scale(%s)", sim)
)
});
names(mods_baseline_window_type) <- sims
lapply(mods_baseline_window_type, function(m) m@optinfo$conv$lme4$messages)$image_similarity
NULL
$text_similarity
NULL
$multimodal_similarity
NULL
$ooo_similarity
NULLlapply(mods_baseline_window_type, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 116735.5
Scaled residuals:
Min 1Q Median 3Q Max
-2.74435 -0.70598 0.06353 0.75144 2.33986
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.005936 0.07705
original_target_label:img_key (Intercept) 0.050519 0.22476
dataset_id (Intercept) 0.001850 0.04301
Residual 0.883599 0.94000
Number of obs: 42614, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) 3.082e-04 1.576e-02 1.805e+01
scale(image_similarity) -1.258e-02 1.361e-02 3.098e+02
scale(age) 1.022e-01 9.172e-03 4.183e+02
scale(aoa) 1.283e-03 9.481e-03 6.035e+02
scale(MeanSaliencyDiff) 1.455e-02 1.029e-02 5.496e+02
scale(trial_window_c) 2.216e-01 4.554e-03 4.142e+04
scale(image_similarity):scale(age) -1.183e-03 7.385e-03 1.723e+03
scale(image_similarity):scale(trial_window_c) -1.011e-02 4.462e-03 4.142e+04
t value Pr(>|t|)
(Intercept) 0.020 0.9846
scale(image_similarity) -0.924 0.3561
scale(age) 11.145 <2e-16 ***
scale(aoa) 0.135 0.8924
scale(MeanSaliencyDiff) 1.414 0.1579
scale(trial_window_c) 48.650 <2e-16 ***
scale(image_similarity):scale(age) -0.160 0.8727
scale(image_similarity):scale(trial_window_c) -2.265 0.0235 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) sc(__) s(_):()
scl(mg_sml) -0.063
scale(age) -0.045 -0.076
scale(aoa) -0.133 -0.096 -0.103
scl(MnSlnD) -0.018 0.089 0.034 0.070
scl(trl_w_) 0.000 0.000 0.000 0.000 0.000
scl(mg_):() -0.072 0.006 -0.318 0.049 -0.050 0.000
scl(_):(__) 0.000 0.000 0.000 0.000 0.000 0.017 0.000
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 116693.4
Scaled residuals:
Min 1Q Median 3Q Max
-2.7779 -0.7042 0.0637 0.7496 2.3743
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.005934 0.07703
original_target_label:img_key (Intercept) 0.050674 0.22511
dataset_id (Intercept) 0.002097 0.04579
Residual 0.882670 0.93951
Number of obs: 42614, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) -7.511e-04 1.656e-02 1.906e+01
scale(text_similarity) -5.940e-03 1.279e-02 1.631e+02
scale(age) 1.033e-01 9.334e-03 5.050e+02
scale(aoa) 6.957e-05 9.811e-03 6.877e+02
scale(MeanSaliencyDiff) 1.541e-02 1.026e-02 5.556e+02
scale(trial_window_c) 2.215e-01 4.551e-03 4.142e+04
scale(text_similarity):scale(age) 6.388e-03 7.790e-03 1.334e+03
scale(text_similarity):scale(trial_window_c) -3.084e-02 4.483e-03 4.142e+04
t value Pr(>|t|)
(Intercept) -0.045 0.964
scale(text_similarity) -0.465 0.643
scale(age) 11.065 < 2e-16 ***
scale(aoa) 0.007 0.994
scale(MeanSaliencyDiff) 1.502 0.134
scale(trial_window_c) 48.665 < 2e-16 ***
scale(text_similarity):scale(age) 0.820 0.412
scale(text_similarity):scale(trial_window_c) -6.879 6.13e-12 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) sc(__) s(_):()
scl(txt_sm) 0.206
scale(age) -0.017 0.076
scale(aoa) -0.065 0.251 -0.037
scl(MnSlnD) -0.005 0.027 0.038 0.089
scl(trl_w_) 0.000 0.000 0.000 0.000 0.000
scl(tx_):() 0.104 -0.049 0.345 0.076 0.034 0.000
scl(_):(__) 0.000 0.000 0.000 0.000 0.000 0.008 0.000
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 116729.9
Scaled residuals:
Min 1Q Median 3Q Max
-2.76169 -0.70593 0.06425 0.75307 2.35370
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.005917 0.07692
original_target_label:img_key (Intercept) 0.050185 0.22402
dataset_id (Intercept) 0.001994 0.04466
Residual 0.883527 0.93996
Number of obs: 42614, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error
(Intercept) -8.625e-04 1.591e-02
scale(multimodal_similarity) -1.314e-02 1.139e-02
scale(age) 1.025e-01 8.781e-03
scale(aoa) -9.827e-04 9.536e-03
scale(MeanSaliencyDiff) 1.498e-02 1.022e-02
scale(trial_window_c) 2.216e-01 4.554e-03
scale(multimodal_similarity):scale(age) 8.754e-03 6.927e-03
scale(multimodal_similarity):scale(trial_window_c) -1.363e-02 4.491e-03
df t value Pr(>|t|)
(Intercept) 1.889e+01 -0.054 0.95735
scale(multimodal_similarity) 5.106e+02 -1.153 0.24937
scale(age) 3.310e+02 11.669 < 2e-16
scale(aoa) 5.980e+02 -0.103 0.91795
scale(MeanSaliencyDiff) 5.518e+02 1.466 0.14322
scale(trial_window_c) 4.142e+04 48.655 < 2e-16
scale(multimodal_similarity):scale(age) 5.186e+03 1.264 0.20633
scale(multimodal_similarity):scale(trial_window_c) 4.142e+04 -3.035 0.00241
(Intercept)
scale(multimodal_similarity)
scale(age) ***
scale(aoa)
scale(MeanSaliencyDiff)
scale(trial_window_c) ***
scale(multimodal_similarity):scale(age)
scale(multimodal_similarity):scale(trial_window_c) **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) sc(__) s(_):()
scl(mltmd_) 0.032
scale(age) -0.075 0.012
scale(aoa) -0.129 0.152 -0.093
scl(MnSlnD) -0.016 0.011 0.024 0.082
scl(trl_w_) 0.000 0.000 0.000 0.000 0.000
scl(ml_):() 0.022 -0.017 0.139 0.012 -0.016 0.000
scl(_):(__) 0.000 0.000 0.000 0.000 0.000 0.013 0.000
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 111349.5
Scaled residuals:
Min 1Q Median 3Q Max
-2.75772 -0.70507 0.06405 0.74717 2.27188
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.006329 0.07955
original_target_label:img_key (Intercept) 0.044911 0.21192
dataset_id (Intercept) 0.001355 0.03681
Residual 0.886115 0.94134
Number of obs: 40622, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) 3.661e-04 1.493e-02 1.511e+01
scale(ooo_similarity) -1.146e-02 1.195e-02 1.276e+02
scale(age) 1.048e-01 8.779e-03 1.942e+02
scale(aoa) -3.265e-03 9.576e-03 4.729e+02
scale(MeanSaliencyDiff) 1.513e-02 1.027e-02 5.085e+02
scale(trial_window_c) 2.224e-01 4.675e-03 3.951e+04
scale(ooo_similarity):scale(age) 1.070e-02 6.707e-03 1.322e+03
scale(ooo_similarity):scale(trial_window_c) -2.848e-02 4.661e-03 3.951e+04
t value Pr(>|t|)
(Intercept) 0.025 0.981
scale(ooo_similarity) -0.959 0.340
scale(age) 11.941 <2e-16 ***
scale(aoa) -0.341 0.733
scale(MeanSaliencyDiff) 1.473 0.141
scale(trial_window_c) 47.571 <2e-16 ***
scale(ooo_similarity):scale(age) 1.596 0.111
scale(ooo_similarity):scale(trial_window_c) -6.111 1e-09 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD) sc(__) s(_):()
scl(_smlrt) 0.114
scale(age) -0.074 -0.012
scale(aoa) -0.159 -0.098 -0.108
scl(MnSlnD) -0.005 0.043 0.023 0.075
scl(trl_w_) 0.000 0.000 0.000 0.000 0.000
scl(_sm):() -0.028 -0.070 0.206 0.009 -0.026 0.000
scl(_):(__) 0.000 0.000 0.000 0.000 0.000 0.044 0.000 I mean these are all interesting, sure.
# 1. Define models and scale constants
mm <- mods_baseline_window_type[[3]]
ooo <- mods_baseline_window_type[[4]]
text <- mods_baseline_window_type[[2]]
image <- mods_baseline_window_type[[1]]
mu <- mean(baseline_data_window_type$target_looking, na.rm = TRUE)
sig <- sd(baseline_data_window_type$target_looking, na.rm = TRUE)
# 2. Generate predictions (using unique variable names)
pred_mm <- ggpredict(mm, terms = c("multimodal_similarity", "trial_window_c [-0.5, 0.5]"))
pred_ooo <- ggpredict(ooo, terms = c("ooo_similarity", "trial_window_c [-0.5, 0.5]"))
pred_text <- ggpredict(text, terms = c("text_similarity", "trial_window_c [-0.5, 0.5]"))
pred_image <- ggpredict(image, terms = c("image_similarity", "trial_window_c [-0.5, 0.5]"))
# 3. Add a source/type label to each dataframe
pred_mm$sim_type <- "Multimodal similarity"
pred_ooo$sim_type <- "Adult behavioral similarity"
pred_text$sim_type <- "Text similarity"
pred_image$sim_type <- "Image similarity"
# 4. Combine all predictions into one dataframe
combined_preds <- rbind(pred_mm, pred_ooo, pred_text, pred_image)
combined_preds$sim_type <- factor(
combined_preds$sim_type,
levels = c(
"Image similarity",
"Text similarity",
"Multimodal similarity",
"Adult behavioral similarity"
)
)
# 5. Un-scale / Back-transform the predicted values and confidence intervals
combined_preds$predicted_raw <- combined_preds$predicted * sig + mu
combined_preds$conf.low_raw <- combined_preds$conf.low * sig + mu
combined_preds$conf.high_raw <- combined_preds$conf.high * sig + mu
# 6. Create the Faceted Plot
baseline_interaction_plot <- ggplot(
combined_preds,
aes(x = x, y = predicted_raw, colour = group, fill = group)
) +
geom_ribbon(
aes(ymin = conf.low_raw, ymax = conf.high_raw),
alpha = 0.15,
colour = NA
) +
geom_line(linewidth = 1.2) +
geom_hline(
yintercept = 0.5,
linetype = "dashed",
alpha = 0.7
) +
# This splits your plot into 4 facets based on similarity type
facet_wrap(~ sim_type, scales = "free_x") +
labs(
title = "Effect of similarities by trial window type",
x = "Target-distractor similarity",
y = "Predicted proportion target looking",
colour = "Trial window",
fill = "Trial window"
) +
scale_color_discrete(
labels = c("Baseline window", "Critical window")
) +
scale_fill_discrete(
labels = c("Baseline window", "Critical window")
) +
theme_classic() +
theme(
strip.background = element_blank(), # Cleans up facet header backgrounds
strip.text = element_text(face = "bold", size = 11), # Emphasizes facet titles
legend.position = "bottom" # Optional: Moves legend to bottom for wider facet viewing
)
ggsave(here("figures/baseline_interaction.png"), baseline_interaction_plot)baseline window looking?
mods_baseline_looking <- lapply(sims, fit_main, data=baseline_data, response="mean_target_looking_baseline_window"); names(mods_baseline_looking) <- sims
lapply(mods_baseline_looking, function(m) m@optinfo$conv$lme4$messages)$image_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$text_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$multimodal_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$ooo_similarity
[1] "boundary (singular) fit: see help('isSingular')"lapply(mods_baseline_looking, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59654.2
Scaled residuals:
Min 1Q Median 3Q Max
-2.01216 -0.82273 -0.00163 0.78419 2.08583
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.000e+00 0.000000
scale(image_similarity) 3.595e-05 0.005996 NaN
original_target_label:img_key (Intercept) 4.611e-02 0.214723
dataset_id (Intercept) 0.000e+00 0.000000
Residual 9.391e-01 0.969055
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) 9.571e-03 1.273e-02 4.461e+02 0.752
scale(image_similarity) 1.300e-03 1.407e-02 4.081e+02 0.092
scale(age) 1.514e-02 1.088e-02 2.486e+03 1.391
scale(aoa) 3.830e-03 1.067e-02 7.540e+02 0.359
scale(MeanSaliencyDiff) 2.862e-02 1.150e-02 5.226e+02 2.488
scale(image_similarity):scale(age) -9.589e-05 9.231e-03 1.947e+03 -0.010
Pr(>|t|)
(Intercept) 0.4524
scale(image_similarity) 0.9264
scale(age) 0.1643
scale(aoa) 0.7197
scale(MeanSaliencyDiff) 0.0132 *
scale(image_similarity):scale(age) 0.9917
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mg_sml) -0.038
scale(age) 0.049 -0.108
scale(aoa) -0.146 -0.110 -0.179
scl(MnSlnD) -0.036 0.069 0.051 0.068
scl(mg_):() -0.132 -0.055 -0.352 0.090 -0.054
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59653.9
Scaled residuals:
Min 1Q Median 3Q Max
-2.01323 -0.82437 -0.00213 0.78454 2.08579
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 1.796e-09 4.238e-05
scale(text_similarity) 1.227e-10 1.108e-05 1.00
original_target_label:img_key (Intercept) 4.616e-02 2.149e-01
dataset_id (Intercept) 1.560e-10 1.249e-05
Residual 9.391e-01 9.691e-01
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) 1.127e-02 1.299e-02 4.351e+02 0.868
scale(text_similarity) -7.033e-04 1.248e-02 4.873e+02 -0.056
scale(age) 1.799e-02 1.100e-02 1.748e+03 1.635
scale(aoa) 4.795e-03 1.111e-02 7.195e+02 0.432
scale(MeanSaliencyDiff) 2.893e-02 1.148e-02 5.248e+02 2.520
scale(text_similarity):scale(age) 6.781e-03 9.551e-03 1.837e+03 0.710
Pr(>|t|)
(Intercept) 0.386
scale(text_similarity) 0.955
scale(age) 0.102
scale(aoa) 0.666
scale(MeanSaliencyDiff) 0.012 *
scale(text_similarity):scale(age) 0.478
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(txt_sm) 0.148
scale(age) 0.090 0.156
scale(aoa) -0.065 0.279 -0.061
scl(MnSlnD) -0.028 0.010 0.060 0.086
scl(tx_):() 0.189 -0.023 0.362 0.124 0.049
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 59641.7
Scaled residuals:
Min 1Q Median 3Q Max
-2.01340 -0.82210 -0.00101 0.78181 2.08068
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.000e+00 0.000e+00
scale(multimodal_similarity) 3.304e-03 5.748e-02
original_target_label:img_key (Intercept) 4.541e-02 2.131e-01
dataset_id (Intercept) 1.279e-10 1.131e-05
Residual 9.356e-01 9.673e-01
Corr
NaN
Number of obs: 21307, groups:
subject_id, 1356; original_target_label:img_key, 865; dataset_id, 24
Fixed effects:
Estimate Std. Error df
(Intercept) 1.018e-02 1.256e-02 4.380e+02
scale(multimodal_similarity) -1.255e-02 1.275e-02 4.578e+02
scale(age) 1.672e-02 1.014e-02 1.843e+03
scale(aoa) 2.798e-03 1.068e-02 7.353e+02
scale(MeanSaliencyDiff) 2.807e-02 1.142e-02 5.250e+02
scale(multimodal_similarity):scale(age) 1.507e-02 9.201e-03 2.139e+03
t value Pr(>|t|)
(Intercept) 0.811 0.4180
scale(multimodal_similarity) -0.984 0.3256
scale(age) 1.650 0.0992 .
scale(aoa) 0.262 0.7935
scale(MeanSaliencyDiff) 2.458 0.0143 *
scale(multimodal_similarity):scale(age) 1.638 0.1016
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mltmd_) 0.010
scale(age) 0.001 0.020
scale(aoa) -0.137 0.153 -0.168
scl(MnSlnD) -0.041 0.012 0.041 0.081
scl(ml_):() 0.044 -0.063 0.097 0.012 -0.026
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 57024.8
Scaled residuals:
Min 1Q Median 3Q Max
-1.98375 -0.82816 -0.00175 0.78846 2.05156
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.000e+00 0.000e+00
scale(ooo_similarity) 1.984e-12 1.409e-06 NaN
original_target_label:img_key (Intercept) 4.111e-02 2.028e-01
dataset_id (Intercept) 0.000e+00 0.000e+00
Residual 9.483e-01 9.738e-01
Number of obs: 20311, groups:
subject_id, 1356; original_target_label:img_key, 795; dataset_id, 24
Fixed effects:
Estimate Std. Error df t value
(Intercept) 1.082e-02 1.266e-02 3.972e+02 0.854
scale(ooo_similarity) -2.591e-04 1.218e-02 4.366e+02 -0.021
scale(age) 1.544e-02 1.023e-02 1.552e+03 1.510
scale(aoa) 5.628e-04 1.079e-02 6.726e+02 0.052
scale(MeanSaliencyDiff) 2.802e-02 1.154e-02 4.776e+02 2.428
scale(ooo_similarity):scale(age) 3.408e-03 8.426e-03 1.816e+03 0.405
Pr(>|t|)
(Intercept) 0.3934
scale(ooo_similarity) 0.9830
scale(age) 0.1313
scale(aoa) 0.9584
scale(MeanSaliencyDiff) 0.0155 *
scale(ooo_similarity):scale(age) 0.6859
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(_smlrt) 0.060
scale(age) 0.002 0.038
scale(aoa) -0.144 -0.101 -0.178
scl(MnSlnD) -0.028 0.015 0.040 0.079
scl(_sm):() 0.018 -0.079 0.131 0.012 -0.019
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')Under 3 years old
LWL specific ages and AoA?
mods_younger <- lapply(sims, fit_main, data=model_data |> filter(age <= 36)); names(mods_younger) <- sims
lapply(mods_younger, function(m) m@optinfo$conv$lme4$messages)$image_similarity
NULL
$text_similarity
[1] "boundary (singular) fit: see help('isSingular')"
$multimodal_similarity
NULL
$ooo_similarity
NULLlapply(mods_younger, function(m) summary(m))$image_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 50134.6
Scaled residuals:
Min 1Q Median 3Q Max
-3.2328 -0.6425 0.1005 0.7465 2.3764
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.042827 0.20695
scale(image_similarity) 0.001451 0.03809 -0.32
original_target_label:img_key (Intercept) 0.051350 0.22660
dataset_id (Intercept) 0.019115 0.13826
Residual 0.883328 0.93986
Number of obs: 18093, groups:
subject_id, 827; original_target_label:img_key, 789; dataset_id, 19
Fixed effects:
Estimate Std. Error df t value
(Intercept) -0.058929 0.037522 17.938544 -1.571
scale(image_similarity) -0.031476 0.017869 371.136722 -1.762
scale(age) 0.107157 0.014508 429.772986 7.386
scale(aoa) -0.008869 0.012023 710.924934 -0.738
scale(MeanSaliencyDiff) 0.011070 0.013234 388.928509 0.836
scale(image_similarity):scale(age) -0.007091 0.011193 768.650303 -0.634
Pr(>|t|)
(Intercept) 0.134
scale(image_similarity) 0.079 .
scale(age) 7.92e-13 ***
scale(aoa) 0.461
scale(MeanSaliencyDiff) 0.403
scale(image_similarity):scale(age) 0.527
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mg_sml) -0.080
scale(age) -0.121 -0.032
scale(aoa) -0.049 -0.089 -0.005
scl(MnSlnD) -0.003 0.110 0.019 0.076
scl(mg_):() -0.066 0.027 -0.071 0.022 -0.025
$text_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 50129.5
Scaled residuals:
Min 1Q Median 3Q Max
-3.2223 -0.6410 0.1006 0.7459 2.3465
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.0420219 0.20499
scale(text_similarity) 0.0008227 0.02868 -1.00
original_target_label:img_key (Intercept) 0.0516830 0.22734
dataset_id (Intercept) 0.0201505 0.14195
Residual 0.8830653 0.93972
Number of obs: 18093, groups:
subject_id, 827; original_target_label:img_key, 789; dataset_id, 19
Fixed effects:
Estimate Std. Error df t value
(Intercept) -6.285e-02 3.864e-02 1.814e+01 -1.626
scale(text_similarity) 2.630e-04 1.790e-02 4.936e+02 0.015
scale(age) 1.160e-01 1.514e-02 4.473e+02 7.659
scale(aoa) -7.581e-03 1.225e-02 6.889e+02 -0.619
scale(MeanSaliencyDiff) 1.338e-02 1.318e-02 3.941e+02 1.016
scale(text_similarity):scale(age) 1.977e-02 1.015e-02 2.727e+03 1.947
Pr(>|t|)
(Intercept) 0.1211
scale(text_similarity) 0.9883
scale(age) 1.17e-13 ***
scale(aoa) 0.5362
scale(MeanSaliencyDiff) 0.3105
scale(text_similarity):scale(age) 0.0516 .
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(txt_sm) 0.147
scale(age) -0.115 -0.012
scale(aoa) -0.021 0.153 0.016
scl(MnSlnD) 0.007 0.019 0.021 0.088
scl(tx_):() 0.016 -0.171 0.234 0.066 0.003
optimizer (nloptwrap) convergence code: 0 (OK)
boundary (singular) fit: see help('isSingular')
$multimodal_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 50133.9
Scaled residuals:
Min 1Q Median 3Q Max
-3.2178 -0.6421 0.0997 0.7466 2.4044
Random effects:
Groups Name Variance Std.Dev.
subject_id (Intercept) 0.04274 0.20672
scale(multimodal_similarity) 0.00376 0.06132
original_target_label:img_key (Intercept) 0.05105 0.22594
dataset_id (Intercept) 0.01995 0.14123
Residual 0.88071 0.93846
Corr
-0.13
Number of obs: 18093, groups:
subject_id, 827; original_target_label:img_key, 789; dataset_id, 19
Fixed effects:
Estimate Std. Error df
(Intercept) -6.735e-02 3.794e-02 1.794e+01
scale(multimodal_similarity) -1.888e-02 1.439e-02 3.775e+02
scale(age) 1.058e-01 1.464e-02 3.895e+02
scale(aoa) -1.261e-02 1.220e-02 7.159e+02
scale(MeanSaliencyDiff) 1.249e-02 1.313e-02 3.942e+02
scale(multimodal_similarity):scale(age) 7.162e-04 9.158e-03 1.540e+03
t value Pr(>|t|)
(Intercept) -1.775 0.0928 .
scale(multimodal_similarity) -1.312 0.1903
scale(age) 7.225 2.66e-12 ***
scale(aoa) -1.034 0.3016
scale(MeanSaliencyDiff) 0.951 0.3421
scale(multimodal_similarity):scale(age) 0.078 0.9377
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(mltmd_) 0.027
scale(age) -0.125 0.011
scale(aoa) -0.046 0.151 0.005
scl(MnSlnD) 0.005 0.038 0.022 0.093
scl(ml_):() 0.019 -0.084 0.106 0.070 0.007
$ooo_similarity
Linear mixed model fit by REML. t-tests use Satterthwaite's method [
lmerModLmerTest]
Formula: f
Data: data
REML criterion at convergence: 47607.9
Scaled residuals:
Min 1Q Median 3Q Max
-3.2332 -0.6411 0.1005 0.7438 2.3779
Random effects:
Groups Name Variance Std.Dev. Corr
subject_id (Intercept) 0.043377 0.20827
scale(ooo_similarity) 0.007083 0.08416 -0.26
original_target_label:img_key (Intercept) 0.048158 0.21945
dataset_id (Intercept) 0.020391 0.14280
Residual 0.873112 0.93440
Number of obs: 17227, groups:
subject_id, 827; original_target_label:img_key, 724; dataset_id, 19
Fixed effects:
Estimate Std. Error df t value
(Intercept) -0.07453 0.03896 17.48204 -1.913
scale(ooo_similarity) -0.03862 0.01829 311.19597 -2.111
scale(age) 0.11482 0.01559 351.56543 7.367
scale(aoa) -0.01224 0.01256 635.24624 -0.975
scale(MeanSaliencyDiff) 0.01439 0.01337 356.41570 1.076
scale(ooo_similarity):scale(age) 0.01566 0.01026 973.00617 1.526
Pr(>|t|)
(Intercept) 0.0723 .
scale(ooo_similarity) 0.0356 *
scale(age) 1.25e-12 ***
scale(aoa) 0.3301
scale(MeanSaliencyDiff) 0.2827
scale(ooo_similarity):scale(age) 0.1272
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Correlation of Fixed Effects:
(Intr) scl(_) scl(g) scal() s(MSD)
scl(_smlrt) 0.138
scale(age) -0.138 -0.055
scale(aoa) -0.075 -0.090 0.004
scl(MnSlnD) 0.019 0.079 0.017 0.076
scl(_sm):() -0.046 -0.214 0.274 0.047 -0.025this makes me think of the cascading age prediction..younger kids: image similarity, older kids: multimodal, even older kids: text only?
Adult behavioral similarity is significant here..AoA effects only appear between 24 and 36 months.
Conference plots
todo: make sure models used are not singular
predict main line – with datasets
unscaled
plot_data <- vanilla |>
pivot_longer(all_of(sims), names_to = "sim_type", values_to = "similarity") |>
mutate(sim_type = factor(sim_type, levels = sims)) # controls panel order
# ---- lmer prediction lines, one stacked data frame ---------------------------
pred <- bind_rows(lapply(sims, function(s) {
m <- mods[[s]]
mf <- model.frame(m) # only the rows lmer kept
# raw similarity range as the model saw it (mf stores the scaled column)
scol <- mf[[grep(paste0("scale\\(", s), names(mf))]]
raw <- as.vector(scol) * attr(scol, "scaled:scale") + attr(scol, "scaled:center")
x <- seq(min(raw), max(raw), length.out = 150)
d <- model_data
nd <- data.frame(x,
age = mean(d$age, na.rm = TRUE),
aoa = mean(d$aoa, na.rm = TRUE),
MeanSaliencyDiff = mean(d$MeanSaliencyDiff, na.rm = TRUE))
names(nd)[1] <- s
pr <- predict(m, newdata = nd, re.form = NA, se.fit = TRUE)
rcol <- mf[[1]]
rc <- attr(rcol, "scaled:center"); rs <- attr(rcol, "scaled:scale")
data.frame(sim_type = s, similarity = x,
y = pr$fit * rs + rc, # un-zscoring
lower = (pr$fit - 1.96 * pr$se.fit) * rs + rc,
upper = (pr$fit + 1.96 * pr$se.fit) * rs + rc)
})) |> mutate(sim_type = factor(sim_type, levels = sims))
library(colorspace)
fig <- ggplot(plot_data, aes(similarity, mean_value,
colour = dataset_name, fill = dataset_name)) +
geom_hline(yintercept = 0.5, linetype = "dashed") +
geom_point(size = 3, alpha = 0.1) +
stat_smooth(method = "glm", geom = "line", alpha = 0.4, linewidth = 1) +
# lmer population line + CI
geom_ribbon(data = pred, inherit.aes = FALSE,
aes(x = similarity, ymin = lower, ymax = upper),
fill = "black", alpha = 0.15) +
geom_line(data = pred, inherit.aes = FALSE,
aes(x = similarity, y = y),
colour = "black", alpha = 0.7, linewidth = 1.2) +
facet_wrap(~ sim_type, nrow = 2, scales = "free_x",
labeller = labeller(sim_type = facet_labs),
strip.position = "top") + # strips read as x-axis titles
scale_y_continuous(breaks = seq(0, 1, 0.2)) +
coord_cartesian(ylim = c(0, 1)) +
scale_color_discrete_qualitative(palette="Set 2") +
#scale_color_manual(values = colorRampPalette(brewer.pal(8, "Dark2"))(24)) +
labs(x = "Target–distractor similarity",
y = "Proportion target looking in critical window",
colour = "Dataset Name", fill = "Dataset Name") +
theme_classic(base_size=16) +
theme(legend.position = "bottom",
#strip.placement = "outside",
strip.text = element_text(size = 20),
panel.spacing = unit(1, "lines"),
plot.margin = margin(t = 5.5, r = 30, b = 5.5, l = 5.5, unit = "pt"))
ggsave(here("figures", "dataset_similarities.png"),
fig, width = 16, height = 13, bg = "white")
fig
scaled
library(ggeffects)
library(purrr)
library(dplyr)
# Grab original scales ONLY for back-transforming the Y axis
y_mean <- mean(model_data_scaled$mean_target_looking_critical_window, na.rm = TRUE)
y_sd <- sd(model_data_scaled$mean_target_looking_critical_window, na.rm = TRUE)
pred_scaled <- map_dfr(sims, function(s) {
res <- ggpredict(mods_scaled[[s]], terms = sprintf("scaled_%s [all]", s))
res |>
as.data.frame() |>
transmute(
sim_type = s,
similarity = x, # Keep X as the Z-score (-2 to +2 range)
y = (predicted * y_sd) + y_mean, # Un-scale Y to 0-1
lower = (conf.low * y_sd) + y_mean, # Un-scale CI lower
upper = (conf.high * y_sd) + y_mean # Un-scale CI upper
)
}) |>
mutate(sim_type = factor(sim_type, levels = sims))scaled with dataset?
clip_data_summarized_scaled <- summarize_similarity_data_collapsed_scaled(model_data_scaled, extra_fields = c("dataset_name"))
clip_data_summarized_scaled_bucket <- summarize_similarity_data_collapsed_scaled(model_data_scaled, extra_fields = c("dataset_name", "age_bucket"))
plot_data_scaled <- clip_data_summarized_scaled |>
rename_with(~ gsub("scaled_", "", .x)) |>
pivot_longer(
cols = all_of(sims),
names_to = "sim_type",
values_to = "similarity"
) |>
mutate(
sim_type = factor(sim_type, levels = sims)
)
fig <- ggplot(plot_data_scaled, aes(similarity, mean_value,
colour = dataset_name, fill = dataset_name)) +
geom_hline(yintercept = 0.5, linetype = "dashed") +
geom_point(size = 3, alpha = 0.1) +
stat_smooth(method = "glm", geom = "line", alpha = 0.5, linewidth = 1) +
# lmer population line + CI
geom_ribbon(data = pred_scaled, inherit.aes = FALSE,
aes(x = similarity, ymin = lower, ymax = upper),
fill = "black", alpha = 0.15) +
geom_line(data = pred_scaled, inherit.aes = FALSE,
aes(x = similarity, y = y),
colour = "black", alpha = 0.7, linewidth = 1.2) +
facet_wrap(~ sim_type, nrow = 2, scales = "free_x",
labeller = labeller(sim_type = facet_labs),
strip.position = "top") +
scale_y_continuous(breaks = seq(0, 1, 0.2)) +
coord_cartesian(ylim = c(0, 1)) +
scale_color_discrete_qualitative(palette="Set 2") +
labs(x = "Target–distractor similarity (z-scored)",
y = "Proportion target looking in critical window",
colour = "Dataset Name", fill = "Dataset Name") +
theme_classic(base_size=16) +
theme(legend.position = "bottom",
strip.text = element_text(size = 20),
panel.spacing = unit(1, "lines"),
plot.margin = margin(t = 5.5, r = 30, b = 5.5, l = 5.5, unit = "pt"))
ggsave(here("figures", "dataset_similarities_zscored.png"),
fig, width = 16, height = 13, bg = "white")
ggsave(here("figures", "dataset_similarities_zscored.svg"),
fig, width = 16, height = 13, device="pdf")age
todo: ensure models used are not singular.
model_data_age <- model_data_scaled |>
filter(!is.na(age_bucket))
### 2. Fit Individual Models using Scaled Data
mods_age <- expand_grid(
sim = sims,
age_bucket = unique(model_data_age$age_bucket)
) |>
mutate(
model = map2(sim, age_bucket, ~
fit_main_scaled(
sim = .x,
data = model_data_age |> filter(age_bucket == .y)
)
),
name = paste(sim, age_bucket, sep = "__")
)
# Convert into a named list of models
mods_age <- setNames(mods_age$model, mods_age$name)
# Note: Ensure 'model_data_age' holds the unscaled target variable,
# or use your original unscaled base data frame here.
y_mean <- mean(model_data_age$mean_target_looking_critical_window, na.rm = TRUE)
y_sd <- sd(model_data_age$mean_target_looking_critical_window, na.rm = TRUE)
# 2. Re-build the prediction grid, handling scaled predictors and un-scaling outcomes
pred <- map_dfr(names(mods_age), function(nm) {
m <- mods_age[[nm]]
parts <- strsplit(nm, "__")[[1]]
st <- parts[1] # Similarity type (sim)
ab <- parts[2] # Age bucket
# Find the scaled similarity term used in this specific model
sim_col <- grep("scaled_.*_similarity$", all.vars(formula(m)), value = TRUE)
if(length(sim_col) == 0) {
# Fallback if your scaled variables don't strictly follow that exact pattern
sim_col <- grep("_similarity$", all.vars(formula(m)), value = TRUE)
}
# Generate predictions based on the z-scored term
ggpredict(m, terms = paste0(sim_col, " [all]")) |>
as.data.frame() |>
transmute(
sim_type = st,
age_bucket = ab,
similarity = x, # Keep X as the Z-score (-2 to +2 range)
y = (predicted * y_sd) + y_mean, # Un-scale Y to 0-1
lower = (conf.low * y_sd) + y_mean, # Un-scale CI lower
upper = (conf.high * y_sd) + y_mean # Un-scale CI upper
)
}) |>
mutate(sim_type = factor(sim_type, levels = sims))# Note: Ensure plot_data_age_scaled has its similarity values z-scored
# but uses the uniform column name 'similarity' after pivoting.
plot_data_scaled_bucket <- clip_data_summarized_scaled_bucket |>
rename_with(~ gsub("scaled_", "", .x)) |>
pivot_longer(
cols = all_of(sims),
names_to = "sim_type",
values_to = "similarity"
) |>
mutate(
sim_type = factor(sim_type, levels = sims)
)
age_scaled_plot <- ggplot(
plot_data_scaled_bucket |> filter(!is.na(age_bucket)),
aes(similarity, mean_value, colour = age_bucket, fill = age_bucket)
) +
geom_hline(yintercept = 0.5, linetype = "dashed") +
geom_point(alpha = 0.12, size = 2) +
# lmer population ribbon + lines split/colored by age_bucket
geom_ribbon(
data = pred |> filter(!is.na(age_bucket)),
aes(x = similarity, ymin = lower, ymax = upper, fill = age_bucket),
inherit.aes = FALSE, alpha = 0.2, colour = NA
) +
geom_line(
data = pred |> filter(!is.na(age_bucket)),
aes(x = similarity, y = y, colour = age_bucket),
inherit.aes = FALSE, linewidth = 1.2
) +
facet_wrap(
~ sim_type,
nrow = 2,
scales = "free_x",
labeller = labeller(sim_type = facet_labs),
strip.position = "top"
) +
scale_y_continuous(breaks = seq(0, 1, 0.2)) +
coord_cartesian(ylim = c(0, 1)) +
scale_color_viridis_d(option = "mako", direction=-1, begin=0.15, end=0.7) +
scale_fill_viridis_d(option="mako", direction=-1, begin=0.15, end=0.7) +
labs(
x = "Target–distractor similarity (z-scored)",
y = "Proportion target looking in critical window",
colour = "Age bucket",
fill = "Age bucket"
) +
theme_classic(base_size = 16) +
theme(
legend.position = "bottom",
strip.text = element_text(size = 20),
panel.spacing = unit(1, "lines"),
plot.margin = margin(t = 5.5, r = 30, b = 5.5, l = 5.5, unit = "pt")
)
ggsave(here("figures", "age_similarities_zscored.png"),
age_scaled_plot, width = 15, height = 13, bg = "white")