In kleinschmidt/phonetic-sup-unsup: Supervised/unsupervised adaptation analysis

What and why

• Want to run follow-up studies that depend on people classifying particular pairs of stimuli as the same or different
• Need to find two conditions shifted by 20ms relative to each other, which maximize the difference in the critical comparisons

library(supunsup)
library(dplyr)
library(magrittr)
library(knitr)
library(ggplot2)

knitr::opts_chunk$set(fig.retina=2,
                      dev='png')

d <- load_and_parse('../../hits/data/supunsup.minusten-hack-visworld.csv')
dd <- supunsup::supunsup_clean %>% filter(supCond == 'unsupervised')

dat <- rbind(d, dd)

GLM-based analysis

Strategy: look at logistic regression fits for each subject, pull out cateogry boundary, and compare.

fits <- dat %>%
  group_by(subject, bvotCond) %>%
  do(fit= glm(respP ~ vot, data=., family='binomial')) %>%
  mutate(int = coef(fit)[1], slope = coef(fit)[2], xint = -int/slope)

ggplot(filter(fits, abs(xint)<100), aes(y=xint, x=bvotCond)) + geom_boxplot() + geom_point(position='jitter') +coord_flip()

Probability same- or different-classification

Look directly at whether or not the relevant pairs are classified as the same or different.

dat %>%
  filter(bvotCond %in% c(-10, 10),
         vot %in% seq(-20, 70, by=20)) %>%
  group_by(subject, bvotCond, vot) %>% 
  summarise(respP = mean(respP)) %>%
  ggplot(aes(x=vot, y=respP, color=bvotCond)) + geom_line(aes(group=subject))


## probability of "same" for pairs of stimuli separated by 20ms
dat %>%
  group_by(bvotCond, subject, vot) %>%
  summarise(respP = mean(respP)) %>%
  mutate(prob_same = lag(respP) * lead(respP) + (1-lag(respP)) * (1-lead(respP)),
         vot_pair = paste(lag(vot), lead(vot))) %>%
  filter(!is.na(prob_same)) %>%
  group_by(bvotCond, vot, vot_pair, subject) %>%
  summarise(prob_same = mean(prob_same)) %>%
  mutate(vot_rel = vot - as.numeric(as.character(bvotCond))) %>%
  filter(mod(vot_rel, 20) == 0) %>%
  ggplot(aes(x=vot, y=prob_same, color=bvotCond)) +
  geom_point(position='jitter') +
  geom_pointrange(stat='summary', fun.data='mean_cl_boot') + 
  facet_grid(.~bvotCond)

There's really not much of a difference between -10 and 20. Looks like the -10/10 pair might have a slight edge over the 0/20. But it's close.

kleinschmidt/phonetic-sup-unsup documentation built on May 20, 2019, 12:33 p.m.