starmans-predictivity.R

library(readxl)
library(tidyverse)
library(caret)
library(brglm2)

df <- read_xlsx("~/../Desktop/190610-Data-All Studies.xlsx") %>%
  filter(StudyNumber == 1) %>%
  dplyr::mutate(
    IsPhilosopher = map_lgl(Field, ~.x == "Philosophy")
  ) %>%
  mutate_at(
    vars(Knowledge, Field, Condition, IsPhilosopher),
    as.factor
  )

# use bias-restricted estimates because rank deficiency
fit <- glm(
  Knowledge ~ Field * Condition,
  data = df,
  family = "binomial",
  method = "brglmFit"
)

broom::tidy(fit, conf.int = TRUE) %>%
  arrange(desc(abs(conf.low))) %>%
  ggplot(aes(term, y = estimate, ymin = conf.low, ymax = conf.high)) +
  geom_errorbar() +
  coord_flip() +
  labs(
    title = "Confidence intervals for Experiment 1 coefficients",
    y = "Logistic regression coefficient"
  ) +
  theme_minimal() +
  theme(
    axis.title.y = element_blank()
  )

ctrl <- trainControl(
  method = "repeatedcv",
  repeats = 10,
  number = 10
)

train(
  Knowledge ~ Field * Condition,
  data = df,
  method = "glm",
  family = "binomial",
  trControl = ctrl
)

train(
  Knowledge ~ IsPhilosopher * Condition,
  data = df,
  method = "glm",
  family = "binomial",
  trControl = ctrl
)