Toward a grammar of behavioural experiments

Danielle Navarro, UNSW (@djnavarro)

Who is she?

A tool to build flexible behavioural experiments in R

Behavioural experiments in the early 20th century

Behavioural experiments in the early 21st century

How does R look to a behavioural scientist?

  • Good statistical analysis
  • Good data wrangling
  • Good general programming
  • Bad for experiments :-(

Python has psychopy

Matlab has psychtoolbox

What about R?

X11 graphics?

X11 graphics?

  # start timing
  start <- Sys.time()
  
  # draw plot
  draw_plot(data$n_up[i], data$n_down[i])
  
  # wait for (then record) response
  data$response[i] <- getGraphicsEvent(
    prompt = "6 = up, b = down",
    consolePrompt = "",
    onKeybd = response
  )
  
  # record the response time
  data$rt[i] <- as.numeric(Sys.time() - start)
  
  # clear the plot
  clear_plot()

X11 graphics?

  • Doesn't work online
  • Interactivity is limited
  • R doesn't "want" you to do it
  • Surprisingly easy to code
  • Few dependencies

Shiny?

https://djnavarro.shinyapps.io/shiny/

Shiny?

https://psyr.org/shiny.html

Shiny?

  • Scaling to MTurk load is hard
  • Tricky for novices
  • Not tested for high-precision timing
  • Intensive at server side
  • Works online
  • Flexible interactivity

???

???

???

???

???

???

Desiderata?

  • Written with R code
  • Relatively easy for novices
  • Experiments are brower based
  • Client side computation
  • Randomisation and flow control
  • Support richer designs than surveys

jsPsych almost works!

  • Experiments are brower based
  • Client side computation
  • Randomisation and flow control
  • Support richer designs than surveys
  • Relatively easy for novices
  • ... But it's javascript

Code the experiment in R with jaysire

Build to jsPsych and R server

Deploy locally or to Google

using the package

(... in brief!)

library(jaysire)

instructions <- trial_instructions(
  pages = c(
    "Welcome! Use the arrow buttons to browse these instructions",
    "Your task is to decide if an equation like '2 + 2 = 4' is true or false",
    "You will respond by clicking a button",
    "Press the 'Next' button to begin!"
  ),
  show_clickable_nav = TRUE,
  post_trial_gap = 1000
)
library(jaysire)

instructions <- trial_instructions(
  pages = c(
    "Welcome! Use the arrow buttons to browse these instructions",
    "Your task is to decide if an equation like '2 + 2 = 4' is true or false",
    "You will respond by clicking a button",
    "Press the 'Next' button to begin!"
  ),
  show_clickable_nav = TRUE,
  post_trial_gap = 1000
)

trial1 <- trial_html_button_response(
  stimulus = "13 + 23 = 36",
  choices = c("true", "false"),
  post_trial_gap = 1000
)

trial2 <- trial_html_button_response(
  stimulus = "17 - 9 = 6",
  choices = c("true", "false")
)
library(jaysire)

instructions <- trial_instructions(
  pages = c(
    "Welcome! Use the arrow buttons to browse these instructions",
    "Your task is to decide if an equation like '2 + 2 = 4' is true or false",
    "You will respond by clicking a button",
    "Press the 'Next' button to begin!"
  ),
  show_clickable_nav = TRUE,
  post_trial_gap = 1000
)

trial1 <- trial_html_button_response(
  stimulus = "13 + 23 = 36",
  choices = c("true", "false"),
  post_trial_gap = 1000
)

trial2 <- trial_html_button_response(
  stimulus = "17 - 9 = 6",
  choices = c("true", "false")
)

build_experiment(
  timeline = build_timeline(instructions, trial1, trial2),
  path = "path_to_folder",
  on_finish = fn_save_locally()
)
image_trial <- trial_image_button_response(
  stimulus = insert_resource("heart.png"), 
  stimulus_height = 400,
  stimulus_width = 400,
  choices = c("Unpleasant", "Neutral", "Pleasant") 
)

video_trial <- trial_video_button_response(
  sources = insert_resource(c("heart.mpg", "heart.webm")), 
  choices = c("Unpleasant", "Neutral", "Pleasant"), 
  trial_ends_after_video = FALSE,
  response_ends_trial = TRUE
)

build_experiment(
  timeline = build_timeline(image_trial, video_trial, audio_trial),
  path = temporary_folder(), 
  resources = build_resources("path_to_resource_folder"),
  use_webaudio = FALSE,
  on_finish = fn_save_locally()
)

Use media files

  • To images:
  • To HTML:
  • To video:
  • To audio:

Respond with a key press...

Etc...

Change response method

Respond with a slider bar...

  • To images:
  • To HTML:
  • To video:
  • To audio:

confidence_scale <- c(
  "Very unconfident", 
  "Somewhat unconfident", 
  "Somewhat confident", 
  "Very confident"
)

page3 <- trial_survey_likert(
  preamble = "How confident in you R skills?",
  questions = list(
    question_likert("Data wrangling?", confidence_scale),
    question_likert("Data visualisation?", confidence_scale),
    question_likert("Statistical modelling?", confidence_scale),
    question_likert("Designing experiments?", confidence_scale),
    question_likert("R markdown documents?", confidence_scale)
  )
)

Create survey pages

Data written to CSV

when deployed to Google App Engine responses are written to the Google Datastore

... when deployed locally

flow control

# ------ everyone should see this ------
page1 <- trial_html_button_response(
  "Do you identify as LGBTIQ+?",
  c("Yes", "No", "Prefer not to say")
)

# ------ only some people should see this! ------
followup <- trial_survey_multi_select(
  question_multi(
    prompt = "Select all that apply",
    options = c(
      "Lesbian", "Gay", "Bisexual/Pansexual", "Transgender", 
      "Nonbinary", "Genderqueer", "Intersex", "Asexual", "Other"
    )
  )
)

Conditional execution

# ------ everyone should see this ------
page1 <- trial_html_button_response(
  "Do you identify as LGBTIQ+?",
  c("Yes", "No", "Prefer not to say")
)

# ------ only some people should see this! ------
followup <- trial_survey_multi_select(
  question_multi(
    prompt = "Select all that apply",
    options = c(
      "Lesbian", "Gay", "Bisexual/Pansexual", "Transgender", 
      "Nonbinary", "Genderqueer", "Intersex", "Asexual", "Other"
    )
  )
)

# ------ a JS function to be evaluated at run time ------
condition_check <- fn_data_condition(button_pressed == "0")

# ------ a conditional timeline ------
page1a <- build_timeline(followup) %>% 
  tl_display_if(condition_check)

Conditional execution

# ------ everyone should see this ------
page1 <- trial_html_button_response(
  "Do you identify as LGBTIQ+?",
  c("Yes", "No", "Prefer not to say")
)

# ------ only some people should see this! ------
followup <- trial_survey_multi_select(
  question_multi(
    prompt = "Select all that apply",
    options = c(
      "Lesbian", "Gay", "Bisexual/Pansexual", "Transgender", 
      "Nonbinary", "Genderqueer", "Intersex", "Asexual", "Other"
    )
  )
)

# ------ a JS function to be evaluated at run time ------
condition_check <- fn_data_condition(button_pressed == "0")

# ------ a conditional timeline ------
page1a <- build_timeline(followup) %>% 
  tl_display_if(condition_check)

Conditional execution

# ------ question to be repeated until answered correctly ------
query <- trial_image_button_response(
  stimulus = insert_resource("heart.png"), 
  stimulus_height = 400,
  stimulus_width = 400,
  choices = c("Unpleasant", "Neutral", "Pleasant"),
  prompt = "You will not be allowed to continue unless you select 'Pleasant'"
)

Looped execution

# ------ question to be repeated until answered correctly ------
query <- trial_image_button_response(
  stimulus = insert_resource("heart.png"), 
  stimulus_height = 400,
  stimulus_width = 400,
  choices = c("Unpleasant", "Neutral", "Pleasant"),
  prompt = "You will not be allowed to continue unless you select 'Pleasant'"
)

# ------ a looped timeline ------
page2 <- build_timeline(query) %>%
  tl_display_while(fn_data_condition(button_pressed != "2"))

Looped execution

why jaysire still sucks...

it's a work in progress

  • the API is clunky
  • it should expose more of the jsPsych API
  • the use of JS callback is limited
  • ... it's a long to-do list

a deeper issue

"trials" are not primitives

  • cloze trials... very specific to linguistics
  • iat trials... very specific to social psychology
  • random dot motion tasks... maybe for monkeys?
  • vsl animate-occlusion trials... specific to Richard Aslin
  • vsl grid-scene trials... specific to Richard Aslin
  • etc...

For anything unusual you need to write a new jsPsych plugin

The R code is used to specify a "trial" by building it from smaller parts

????????

This code is used to construct task-specific jspsych plugins at build time

data +

aesthetics +

layers +

scales +

coordinate +

theme

decomposing plots

??? +

??? +

??? +

??? +

??? +

???  

decomposing experiments?

????????

what might it look like?

????????

dx.doi.org/10.3758/s13414-011-0131-9

what might it look like?

Elements:

  • image
  • audio
  • video
  • HTML

 

????????

Stimulus:

  • comprised of elements
  • has spatial layout
  • has temporal layout

what might it look like?

Responses:

  • html button
  • key press
  • mouse move
  • slider
  • check box
  • text box

 

????????

Events:

  • record data
  • modify stimulus
  • provide feedback
  • end trial

Feedback:

  • specialised type of stimulus???

A tool to build flexible behavioural experiments in R

code

build

deploy

templates & randomisation

# ----------- template -----------
trial_template <- trial_html_button_response(
  stimulus = insert_variable("stimulus"),
  choices = c("true", "false"),
  post_trial_gap = 1000
)

use templates & variables...

# ----------- template -----------
trial_template <- trial_html_button_response(
  stimulus = insert_variable("stimulus"),
  choices = c("true", "false"),
  post_trial_gap = 1000
)

# ----------- timeline -----------
trials <- build_timeline(trial_template) %>%
  tl_add_variables(stimulus = c("13 + 23 = 36", "17 - 9 = 6", "etc")) %>%
  tl_add_parameters(randomize_order = TRUE, repetitions = 2)

... to randomise & repeat

other options?

Shiny?

  • Works online
  • Flexible interactivity
  • Simpler with psychTestR
  • Still inherits some of the other issues :-(

formr?

formr.org looks awesome, but targeted at surveys

a demo experiment

Toward a grammar for behavioural experiments

By Danielle Navarro

Toward a grammar for behavioural experiments

Invited talk for rstudio::conf(2020L)

  • 1,772
Loading comments...

More from Danielle Navarro