Spoken word recognition of children with cochlear implants [BUCLD42]
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Spoken word recognition of children with cochlear implants

Research compendium for our presentation at BUCLD 42:

Mahr, T., & Edwards, J. R. (2017, November). Spoken word recognition of children with cochlear implants. Presentation at the 42nd annual Boston University Conference on Language Development, Boston, MA.

Repository contents

Analysis notebooks

00-get-data.Rmd (output) downloads the raw data from our lab's internal database. It cannot be reproduced by people outside our lab. Therefore, I try to keep the script brief and have it save the needed data into the folder data-raw. Later processing steps can be reproduced in the following notebooks.

01-data-screening.Rmd (output) screens the raw eyetracking data, removing trials and participants with excessive missing data.

02-plot-data.Rmd (output) includes various plots of the eyetracking data from the experiment. We had equivocated about how we would go about revealing the plots through the presentation so a lot of options are covered. This file also computes descriptives stats and prepares by-group plots of test scores.

03-models.Rmd (output) runs growth curve analyses for the data.

04-maybe-bias.Rmd (output) runs an additional exploratory growth curve model where trials are separated based on which image the child fixates on at noun onset (the trial's bias).


data-raw/: Raw data downloaded from the database.

data/: Screened data ready for plotting or modeling.

Large csv files are saved as compressed .csv.gz files. The readr package, specifically the function readr::read_csv(), can automatically uncompress these files.

assets/: The image files and model caches for the notebooks are stored here.


mahr-edwards-bucld42-ci-word-recognition.pdf contains the Power Point slides as a pdf file.

plotting-helpers.R stores helpers functions that are recycled in the different scripts.

Makefile is a makefile to automate running and rendering the notebook files.

The code's LICENSE is the GPL-3, but I don't think that should matter because the code is really tailored for this data-set. The data is copyrighted and belongs to the University of Wisconsin–Madison, I think. (We collected data at two different university labs.) I don't know.

Additional information

Reproducibility tips

Don't hesitate to ask here with an issue or by emailing Tristan. These things can be finicky on other people's computers.

This code should install most of the needed packages. Otherwise, the session-info at the bottom of each script has the version and provenance of all R packages used.

install.packages(c("knitr", "rmarkdown", "tidyverse", "lme4", 
                 "yaml", "polypoly", "devtools", "rprojroot",

The easiest way to reproduce the analysis is to clone the repository from GitHub into an RStudio project. In RStudio: File > New Project > Version Control > Git > paste in the URL of this repository. You can "Knit" the individual Rmd files to run and compile the notebooks, or you can use the Build tab to run the Makefile which updates any files that need to be updated.

Abstract from conference handbook

Spoken word recognition of children with cochlear implants

Tristan Mahr (University of Wisconsin–Madison)
Jan Edwards (University of Maryland)

Children with cochlear implants (CIs) perform more poorly than normal hearing (NH) children on virtually every aspect of speech and language. While these deficits have been attributed in large part to the impoverished signal, children with CIs may also have different processing strategies because of their consistent experience of listening to an impoverished signal. We used the mispronunciation paradigm to examine processing differences in word recognition. 26 children with CIs (age: 31–66 months) were compared to NH children matched for age and sex. Children with CIs recognized one-feature mispronunciations as reliably as their NH peers, but it took them longer to look away from the familiar image. They were also slower and less reliable at recognizing highly familiar words, even in a two-image LWL paradigm. These results suggest that the impoverished signal of the CI may result in more uncertainty and longer lexical decision times, especially for ambiguous speech cues.