add mp-norming closed-set responses

[tjmahr]

Child did better on the nonwords than the mispronunciations. Something something Swingley and Aslin (2007).

library(L2TDatabase)
library(dplyr)
library(ggplot2)

cnf_file <- file.path(getwd(), "inst/l2t_db.cnf")
l2t <- l2t_connect(cnf_file, db_name = "l2t")

results <- tbl(l2t, "MPNormingClosed_Items") %>% 
  collect() %>% 
  rename_all(stringr::str_replace, "MPNormingClosed_", "")

ggplot(results) + 
  aes(x = Type, y = Correct) + 
  stat_summary(
    aes(group = ResearchID), 
    position = position_jitter(width = .3, height = 0.02),
    color = "grey20") + 
  stat_summary(
    aes(color = Type))

Okay, here are the participants' abilities as estimated in a Bayesian mixed effects item response model.

library(rstanarm)

m <- stan_glmer(
  Correct ~ Type + (1 | ResearchID/Type) + (1 | Item),
  family = binomial,
  data = results,
  prior = normal(0, 1))

# Get one row per ID and condition so the prediction function
# can condition on the ID and ID:Type effects but not the Item
# effects
try <- results %>% 
  distinct(ResearchID, Type, Item, ItemNumber) %>% 
  group_by(ResearchID, Type) %>% 
  top_n(1, ItemNumber) %>% 
  ungroup()

# This is a helper from my personal Bayesian package :)
p <- tristan::augment_posterior_linpred(
  m, 
  newdata = try, 
  re.form = ~ (1 | ResearchID/Type), 
  transform = TRUE)

p_mean <- tristan::augment_posterior_linpred(
  m, 
  newdata = try %>% filter(ResearchID == "001L"), 
  re.form = NA, 
  transform = TRUE)

ggplot(p) + 
  aes(x = Type, y = .posterior_value) + 
  stat_summary(
    aes(group = ResearchID), 
    fun.data = median_hilow,
    fun.args = list(conf.int = .9),
    position = position_jitter(width = .25),
    alpha = .4) + 
  stat_summary(
    aes(color = Type),
    data = p_mean,
    fun.data = median_hilow,
    fun.args = list(conf.int = .9),
    size = 1) + 
  guides(color = FALSE) +
  labs(
    x = NULL, 
    y = "Proportion correct", 
    caption = "Points: Posterior median and 90% intervals")