parameter_glm: Quantify and plot associations between learning parameters...

In qdercon/pstpipeline: Probabilistic Selection Task (PST) pipeline: parsing, analysis, simulation, and plotting

Quantify and plot associations between learning parameters and variables of interest

Description

parameter_glm is a wrapper around cmdstan_glm(), taking as inputs summary tables and raw data from fit_learning_model(), and outputting the results of GLMs quantifying the association between the individual-level posterior means of each parameter and the independent variable(s) of interest. Gamma GLMs with log link functions are used for learning rate and decay factor models (i.e., positively skewed and constrained between 0 and 1) while standard Gaussian models with identity link functions are used for models with inverse temperature or weights as the response variable.

Usage

parameter_glm(
  summary_df = list(),
  raw_df = list(),
  var_of_interest,
  covariates,
  affect_number = NULL,
  interaction = NULL,
  recode_na = NULL,
  extra_data = NULL,
  rhat_upper = 1.1,
  ess_lower = 100,
  ...
)

Arguments

summary_df	List of cmdstanr::summary() outputs for the fit(s) of interest.
raw_df	List of raw data inputs to the above fits (in the same order). Used to correctly link subject IDs to independent variables.
var_of_interest	Variable of interest.
covariates	Vector of covariates to control for in the GLMs.
affect_number	For affect model fits, specify the number (i.e., 1, 2, or 3) of the affect noun/verb of interest. If affect parameters are found in model summaries, and this is not specified, GLMs will default to standard Q-learning parameters.
interaction	Optional variable to interact with the variable of interest. The GLMs will then be run twice with this variable reverse coded the second time to obtain posterior samples for the variable of interest in both groups. This variable must be binary and only 1 interaction is allowed.
recode_na	Some demographic questions were conditional, and so there exist NAs. This argument allows these terms to be recoded as appropriate (in all binary cases, this should be set to 0).
extra_data	Option to supply a data frame with additional derived quantities (e.g., factor scores). Must include a subjID column.
rhat_upper, ess_lower	Same as plot_raincloud().
...	Other arguments to pass to cmdstan_glm() (e.g., to control number of warm-up and sampling iterations). In addition, use cores to change the number of parallel chains to sample from.

Value

A posterior::draws_df().

Examples

## Not run: 
# Comparing parameters across groups

data(example_data)

fit_nd <- fit_learning_model(
  example_data$nd,
  model = "2a",
  vb = FALSE,
  exp_part = "training"
)
fit_dis <- fit_learning_model(
  example_data$dis,
  model = "2a",
  vb = FALSE,
  exp_part = "training"
)

distanced <- parameter_glm(
  summary_df = list(fit_nd$summary, fit_dis$summary),
  raw_df = list(fit_nd$raw_df, fit_dis$raw_df),
  var_of_interest = "distanced",
  covariates = c("age", "sex", "digit_span"),
  iter_warmup = 1000, iter_sampling = 1000
)

# Comparing affect model parameters w.r.t. anxiety/depression factor scores
# with interaction on distancing

factor_scores <- read.csv("data-raw/gillan_scores.csv")[-1] # from Github

fit_affect_nd <- fit_learning_model(
  example_data$nd,
  model = "2a",
  affect = TRUE,
  exp_part = "training"
)
fit_affect_dis <- fit_learning_model(
  example_data$dis,
  model = "2a",
  affect = TRUE,
  exp_part = "training"
)

AD_affect_all <- pstpipeline::parameter_glm(
  summary_df = list(fit_affect_nd$summary, fit_affect_dis$summary),
  raw_df = list(fit_affect_nd$raw_df, fit_affect_dis$raw_df),
  var_of_interest = "AD",
  covariates = c("age", "sex", "digit_span"),
  interaction = "distanced",
  affect_number = 1,
  extra_data = factor_scores
)

## End(Not run)

qdercon/pstpipeline documentation built on June 1, 2025, 1:11 p.m.