A quick reference to GDINA R package

What can this package do?

This package provides a framework for conducting a series of cognitive diagnosis analysis for dichotomous and polytomous responses.

For more details, see here

For an example of how to use the package, see here

How to use this package?

To estimate various CDMs, use GDINA() function
To extract item and person parameters from GDINA estimates, use coef() and personparm() functions, respectively
To extract other estimation components, use extract() function
To evaluate absolute model fit, use modelfit(), itemfit() or itemfitPD() function
To evaluate relative model fit, use AIC(), BIC() or deviance() function
To validate Q-matrix, use Qval() function
To detect differential item functioning, use dif() function
To regress estimated attribute or profile classifications on covariates with three-step ML correction, use ThreeStepCov()
To run graphical user interface, use startGDINA() function
To simulate data from various CDMs, use simGDINA() function

Visit here for an online reference

A quick example

Below is an illustration showing some analyses that you could do using the package. You can copy the code, paste and run it in R console. For a more comprehensive example, see this tutorial.

Load the package and assume we have the following item response matrix and Q-matrix:

library(GDINA)
dat <- sim10GDINA$simdat
Q <- matrix(c(1,0,0,
              0,1,0,
              0,0,1,
              1,0,1,
              0,1,1,
              1,1,0,
              0,0,1,
              1,0,0,
              1,1,1,
              1,0,1),byrow = T,ncol = 3)

Estimate the G-DINA model:

est <- GDINA(dat = dat, Q = Q, model = "GDINA")

Conduct the Q-matrix validation. By default, it implements de la Torre and Chiu’s (2016) algorithm using a fixed cutoff. This is fast but sometimes suggests too many modifications.

Qv <- Qval(est)
Qv

To avoid using fixed cutoffs and also take uncertainty in item parameter estimation into account, you may consider the stepwise method:

Qv2 <- Qval(est,method = "Wald")
Qv2

To further examine the q-vectors that are suggested to be modified, you can draw mesa plots:

plot(Qv, item = 9)

Perform item-level model selection to see if the saturated G-DINA model can be simplified:

mc <- modelcomp(est)
mc

Assess model-data fit at test and item levels:

# test level absolute fit
mft <- modelfit(est)
mft
# item level absolute fit
ift <- itemfit(est)
ift
summary(ift)
plot(ift)

Calculate classification accuracy

CA(est)

Regress estimated classifications on user covariates using the three-step ML correction.

The bundled real-data examples do not include respondent-level covariates, so in the code below the covariates are just placeholders for user-supplied background variables collected outside the response matrix. The GDINA fit is real, based on the ECPE data, while the covariate frame Z is only used to illustrate the ThreeStepCov() workflow.

dat <- realdata_ECPE$dat
Q <- realdata_ECPE$Q
ecpe_fit <- GDINA(dat, Q, model = "GDINA")

set.seed(123)
N <- nrow(dat)
Z <- data.frame(
  age = round(rnorm(N, mean = 21, sd = 3)),
  female = rbinom(N, 1, 0.55),
  program = factor(sample(c("intensive", "regular", "self-study"), N, replace = TRUE))
)

# Attribute-level regression with the three-step correction
att_reg <- ThreeStepCov(
  ecpe_fit,
  ~ age + female + program,
  data = Z,
  level = "attribute"
)
att_reg$results[[1]]$corrected$table
att_reg$results[[1]]$naive$table

# Profile-level regression for the estimated latent classes
profile_reg <- ThreeStepCov(
  ecpe_fit,
  ~ age + female + program,
  data = Z,
  level = "profile"
)
head(profile_reg$results$corrected$table)

At the attribute level, each element of att_reg$results contains both the naive logistic regression and the corrected regression for one attribute. At the profile level, profile_reg$results$corrected$table returns the multinomial logit coefficients relative to the reference latent profile.

More Examples

If you would like to contribute an example to this website, please send me your .Rmd file.