mirt: RMSD_DIF – R documentation

Become an expert in R — Interactive courses, Cheat Sheets, certificates and more!

RMSD_DIF

RMSD effect size statistic to quantify category-level DIF

Description

This function computes a set of RMSD "badness-of-fit" statistics when investing DIF across a set of grouping variables. In a first step, a (potentially highly constrained) multiple group model is fitted, while in a second step the item (and person) parameters are estimated based on all examines across all groups. Category level DIF is assessed based on how well the pseudo-table of counts match the (constrained) probability functions implied by the original multiple group model (while also weighing across the implied density function of the latent traits). If the RSMD fit is poor, indicating non-ignorable DIF, then the multiple-group model should be adjusted to better account for the large response bias due to using a pooled model. See Lee and von Davier (2020) and Buchholz and Hartig (2019) for details.

Usage

RMSD_DIF(pooled_mod, flag = 0, probfun = TRUE, dentype = "norm")

Arguments

`pooled_mod`	a multiple-group model (used to compute the model-implied probability in the goodness-of-fit test)
`flag`	a numeric value used as a cut-off to help flag larger RMSD values (e.g., `flag = .03` will highlight only categories with RMSD values greater than .03)
`probfun`	logical; use probability functions to compute RMSD? If FALSE, the expected score functions will be integrated instead, which may be useful for collapsing across the categories in polytomous items
`dentype`	density to use for the latent trait. Can be `'norm'` to use a normal Gaussian density where the mean/variance are extracted from the model object(default), `'snorm'` for a standard normal distribution, or `'empirical'` to use the density estimate obtained via the E-table

Author(s)

Phil Chalmers rphilip.chalmers@gmail.com

References

Buchholz, J., and Hartig, J. (2019). Comparing Attitudes Across Groups: An IRT-Based Item-Fit Statistic for the Analysis of Measurement Invariance. Applied Psychological Measurement, 43(3), 241-250.

Lee, S. S., and von Davier, M. (2020). Improving measurement properties of the PISA home possessions scale through partial invariance modeling. Psychological test and assessment modeling, 62(1):55-83.

Examples

## Not run: 

#----- generate some data
set.seed(12345)
a <- a2 <- matrix(abs(rnorm(15,1,.3)), ncol=1)
d <- d2 <- matrix(rnorm(15,0,.7),ncol=1)

# item 1 has DIF
d2[1] <- d[1] - .5
a2[1] <- a[1] + 1

itemtype <- rep('2PL', nrow(a))
N <- 1000
dataset1 <- simdata(a, d, N, itemtype)
dataset2 <- simdata(a2, d2, N, itemtype)
dat <- rbind(dataset1, dataset2)
group <- c(rep('D1', N), rep('D2', N))

#-----

# fully pooled model
pooled_mod <- multipleGroup(dat, 1, group=group,
   invariance = c(colnames(dat), 'free_mean', 'free_var'))
coef(pooled_mod, simplify=TRUE)

RMSD_DIF(pooled_mod)
RMSD_DIF(pooled_mod, dentype = 'empirical')
RMSD_DIF(pooled_mod, flag = .03)

# more freely estimated model (item 1 has 2 parameters estimated)
MGmod <- multipleGroup(dat, 1, group=group,
                       invariance = c(colnames(dat)[-1], 'free_mean', 'free_var'))
coef(MGmod, simplify=TRUE)

# RMSD in item.1 now reduced (MG model accounts for DIF)
RMSD_DIF(MGmod)
RMSD_DIF(MGmod, flag = .03)


#################
# polytomous example
set.seed(12345)
a <- a2 <- matrix(rlnorm(20,.2,.3))

# for the graded model, ensure that there is enough space between the intercepts,
# otherwise closer categories will not be selected often (minimum distance of 0.3 here)
diffs <- t(apply(matrix(runif(20*4, .3, 1), 20), 1, cumsum))
diffs <- -(diffs - rowMeans(diffs))
d <- d2 <- diffs + rnorm(20)

# item 1 has slope + dif for first intercept parameter
d2[1] <- d[1] - .5
a2[1] <- a[1] + 1

itemtype <- rep('graded', nrow(a))
N <- 1000
dataset1 <- simdata(a, d, N, itemtype)
dataset2 <- simdata(a2, d2, N, itemtype)
dat <- rbind(dataset1, dataset2)
group <- c(rep('D1', N), rep('D2', N))

#-----

# fully pooled model
pooled_mod <- multipleGroup(dat, 1, group=group,
         invariance = c(colnames(dat), 'free_mean', 'free_var'))
coef(pooled_mod, simplify=TRUE)

# Item_1 fits poorly in several categories (RMSD > .05)
RMSD_DIF(pooled_mod)
RMSD_DIF(pooled_mod, flag = .05)
RMSD_DIF(pooled_mod, flag = .1, probfun = FALSE) # use expected score function

# more freely estimated model (item 1 has more parameters estimated)
MGmod <- multipleGroup(dat, 1, group=group,
                       invariance = c(colnames(dat)[-1], 'free_mean', 'free_var'))
coef(MGmod, simplify=TRUE)

# RMSDs in Item_1 now reduced (MG model better accounts for DIF)
RMSD_DIF(MGmod)
RMSD_DIF(MGmod, flag = .05)
RMSD_DIF(MGmod, probfun = FALSE, flag = .1) # use expected score function


## End(Not run)

mirt

Multidimensional Item Response Theory

v1.33.2

GPL (>= 3)

Authors

Phil Chalmers [aut, cre] (<https://orcid.org/0000-0001-5332-2810>), Joshua Pritikin [ctb], Alexander Robitzsch [ctb], Mateusz Zoltak [ctb], KwonHyun Kim [ctb], Carl F. Falk [ctb], Adam Meade [ctb], Lennart Schneider [ctb], David King [ctb], Chen-Wei Liu [ctb], Ogreden Oguzhan [ctb]

Initial release