Scribd
Upload
44 views

GuideToIRTinvarianceUsingMIRT (ANCHOR)

This document provides guidance on testing measurement invariance using the MIRT package in R. It outlines the following steps: 1) Testing assumptions like dimensionality and model fit; 2) Creating a constrained baseline model; 3) Conducting a first round of likelihood ratio tests with all-others-as-anchors; 4) Specifying a new baseline model using anchor items; 5) Running a final round of invariance tests. Functions are provided to help identify differentially and non-differentially functioning items. The guidance is illustrated with sample data containing inserted differential item functioning.

Uploaded by

Ampeh Sarok
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
44 views

GuideToIRTinvarianceUsingMIRT (ANCHOR)

This document provides guidance on testing measurement invariance using the MIRT package in R. It outlines the following steps: 1) Testing assumptions like dimensionality and model fit; 2) Creating a constrained baseline model; 3) Conducting a first round of likelihood ratio tests with all-others-as-anchors; 4) Specifying a new baseline model using anchor items; 5) Running a final round of invariance tests. Functions are provided to help identify differentially and non-differentially functioning items. The guidance is illustrated with sample data containing inserted differential item functioning.

Uploaded by

Ampeh Sarok
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

1

Guide to IRT Invariance Using the MIRT


Package in R
Adam W. Meade

North Carolina State University

Current as of April 28, 2016

Contents
Getting Started.............................................................................................................................................. 2
Step 1: Testing Assumptions ......................................................................................................................... 5
Parallel analysis ......................................................................................................................................... 5
Model Fit ................................................................................................................................................... 5
Check for Sufficient Number of Category Responses ............................................................................... 5
Create a Constrained Baseline Model........................................................................................................... 6
First Round of LRTs ....................................................................................................................................... 7
Specify a New Baseline Model using Anchor Items ...................................................................................... 7
Run the Final Invariance Tests ...................................................................................................................... 8
Compute Effect Sizes .................................................................................................................................... 9

This document assumes the use of the current version of both R and R Studio. See
https://www.rstudio.com/.
2

Getting Started
To get started, there are several packages required to use this guide. If these packages are not already
installed, first run the following code:

#first load some needed libraries


install.packages("psych")
install.packages("lessR")
install.packages("devtools")

library("psych")
library("lessR")
library("devtools")

Currently, it is best to install the developer version of the MIRT package. This will no longer be
necessary in the future.

#https://github.com/philchalmers/mirt
install_github('philchalmers/mirt')
library('mirt')
3

This step is completely optional, but I have written a function to help make it easier to identify
differentially functioning (DF) and non-DF items. Let’s go ahead and load it for later:

######################################################################
# Functions
#######################################################################
get.dif.items <- function(out.list,p.val,parms){
dif.items <- NULL
non.dif.items <- NULL
for(i in 1:length(out.list)){ # loop list of items
chi.sq <- out.list[[i]][2,6] #2 groups, so second row, and 6th column
df <- out.list[[i]][2,7]
p <- out.list[[i]][2,8]
i.name <- names(out.list[i])
d <- c(i.name,chi.sq,df,p,parms[i,])

if(p < p.val){


dif.items <- rbind(dif.items,d)
}else{
non.dif.items <- rbind(non.dif.items,d)
}
}
if (!is.null(dif.items)) {
dif.items <- data.frame(dif.items, row.names = NULL)
colnames(dif.items)[1] <- "item"
colnames(dif.items)[2] <- "chi.sq"
colnames(dif.items)[3] <- "df"
colnames(dif.items)[4] <- "p"

}
if (!is.null(non.dif.items)) {
non.dif.items <- data.frame(non.dif.items, row.names = NULL)
colnames(non.dif.items)[1] <- "item"
colnames(non.dif.items)[2] <- "chi.sq"
colnames(non.dif.items)[3] <- "df"
colnames(non.dif.items)[4] <- "p"
}
r.list <- list(dif_items = dif.items, no_dif = non.dif.items)
return(r.list)
}
########################## END FUNCTIONS ########################
4

Next we will make some data for illustration purposes using a function.
########## Make the data with a lack of invariance ######
make.data <- function(N){
set.seed(1234)
a <- matrix(abs(rnorm(15,1,.3)), ncol=1)
d <- matrix(rnorm(15,0,.7),ncol=1)
d1 <- d2 <- cbind(d, d-1, d-2) # b parameters for both groups
d2[13:15, ] <- d1[13:15, ] + 1 # here is the DIF
itemtype <- rep('graded', nrow(a))
dataset1 <- simdata(a, d1, N, itemtype)
dataset2 <- simdata(a, d2, N, itemtype)
dat <- rbind(dataset1, dataset2)
return(dat)
}
N <- 1000
dat <- make.data(N)
group <- c(rep('Ref', N), rep('Foc', N))
focal.data <- dat[1:1000,]
ref.data <- dat[1001:2000,]

Here we have specified our sample size per group and given the groups labels. In general, you want your
reference group to be the first group.
5

Step 1: Testing Assumptions


Parallel analysis
Start by examining dimensionality. I’m using parallel analysis from the Psych package. I like to do this
for each group separately.
########### check dimensionality ########
fa.parallel(focal.data)
fa.parallel(ref.data)

Looking at the output, it’s clearly unidimensional.

Model Fit
Next, compute stat and examine plots.
########## check model fit ##############
foc.model <- mirt(focal.data, model = 1, itemtype = "graded", SE=TRUE)
M2(foc.model)
ref.model <- mirt(ref.data, model = 1, itemtype = "graded", SE=TRUE)
M2(ref.model)
foc.fit <- itemfit(foc.model)
foc.fit
ref.fit <- itemfit(ref.model)
ref.fit

### optional plotting


plots.foc <- list()
plots.ref <- list()
for(i in 1:length(dat)){
plots.foc[[i]]<-itemfit(foc.model,empirical.plot = i)
plots.ref[[i]]<-itemfit(ref.model,empirical.plot = i)
}
plots.foc
plots.ref

All looks good.

Check for Sufficient Number of Category Responses


With Likert-type scales, often there will be very few respondents endorsing some of the extreme
response options. If you haven’t already done so, consider collapsing response categories where there
are fewer than 20 respondents and/or where the SE associated with the item b parameter is greater
than around .4. Different items can have different numbers of response options, but the number of
response options must be the same across groups. Note – these are unpublished rules of thumb, so use
at your own risk.
6

Create a Constrained Baseline Model


We will use likelihood ratio tests (LRTs). To start, we need to create a baseline model in which all items
have parameters constrained across items. We also will take a look at the parameters in the constrained
model.
################## Baseline Model ############################
model.constrained <- multipleGroup(dat, 1, group,
invariance = c(colnames(dat), 'free_means', 'free_var'))
constrained.parameters <- coef(model.constrained,simplify = TRUE)[[1]][[1]]
constrained.parameters
7

First Round of LRTs


First, test each of the items by freeing the parameters of each item, one at a time. The other items
serve as the anchor items (i.e., the all-others-as-anchors model).
####### First round of DIF analysesb - All Others As Anchors #######
dif.drop <- DIF(model.constrained, c('a1','d1','d2','d3'), scheme = 'drop',
seq_stat = .05)
dif.drop
## use the optional function to table the output
dif.drop.out <-
get.dif.items(out.list=dif.drop,p.val=.05,parms=constrained.parameters)
dif.drop.out

Examine the output:


$dif_items
item chi.sq df p a1 d1 d2 d3
1 Item_1 17.706 4 0.0014 0.688 -0.133 -1.227 -2.189
2 Item_3 11.086 4 0.0256 1.2 -0.742 -1.711 -2.531
3 Item_5 11.569 4 0.0209 1.086 1.496 0.533 -0.504
4 Item_6 17.88 4 0.0013 1.18 -0.113 -1.095 -2.106
5 Item_7 10.227 4 0.0368 0.836 -0.443 -1.49 -2.451
6 Item_11 9.673 4 0.0463 0.739 -1.155 -2.071 -3.185
7 Item_13 84.557 4 0 0.857 -0.268 -1.275 -2.36
8 Item_14 76.865 4 0 1.045 0.3 -0.624 -1.583
9 Item_15 56.403 4 0 1.216 -0.396 -1.334 -2.218

$no_dif
item chi.sq df p a1 d1 d2 d3
1 Item_2 5.308 4 0.2571 1.094 -0.452 -1.527 -2.548
2 Item_4 1.156 4 0.8853 0.231 -0.579 -1.598 -2.739
3 Item_8 5.054 4 0.2818 0.904 -0.471 -1.56 -2.577
4 Item_9 6.721 4 0.1514 0.754 0.269 -0.782 -1.768
5 Item_10 7.881 4 0.096 0.652 -0.563 -1.502 -2.445
6 Item_12 7.985 4 0.0921 0.68 0.333 -0.657 -1.745

Specify a New Baseline Model using Anchor Items


We will use the A5 method from Meade and Wright (2012) in which we will choose five anchor items
with the largest A parameters.
##### Run an anchor-item model #####
itemnames <- colnames(dat)
anc.items.names <- itemnames[c(2,8,9,10,12)]
test.items <- c(1,3:7,11,13:15)
model_anchor <- multipleGroup(dat, model = 1, group = group,
invariance = c(anc.items.names, 'free_means', 'free_var'))
anchor.parms <-coef(model_anchor,simplify = TRUE)[[1]][[1]]
anchor.parms

The above runs the model and also requests the parameters.
8

Run the Final Invariance Tests


#################### Final round of DIF analyses #################
dif.anchor <- DIF(model_anchor, c('a1','d1','d2','d3'), items2test =
test.items, plotdif = TRUE)
dif.anchor
## use the optional function to table the output
dif.anchor.out <-
get.dif.items(out.list=dif.anchor,p.val=.05,parms=anchor.parms)
dif.anchor.out

Check out the output. Looks like the correct items were detected.
> dif.anchor.out
$dif_items
item chi.sq df p a1 d1 d2 d3
1 Item_13 94.622 4 0 0.753 0.342 -0.71 -1.697
2 Item_14 84.691 4 0 0.656 -0.502 -1.442 -2.386
3 Item_15 64.678 4 0 0.719 -1.018 -1.939 -2.953

$no_dif
item chi.sq df p a1 d1 d2 d3
1 Item_1 7.514 4 0.1111 0.737 -0.01 -1.045 -1.976
2 Item_3 2.072 4 0.7225 1.288 -0.605 -1.655 -2.488
3 Item_4 1.114 4 0.892 0.274 -0.562 -1.617 -2.75
4 Item_5 3.916 4 0.4175 1.204 1.645 0.702 -0.426
5 Item_6 2.661 4 0.6161 1.287 0.067 -0.968 -2.036
6 Item_7 2.484 4 0.6475 0.832 -0.325 -1.366 -2.262
7 Item_11 3.511 4 0.4762 0.91 -0.386 -1.477 -2.496
9

Compute Effect Sizes


The last step is to compute effect size estimates, as described in Meade (2010).
################## Compute the effect sizes #####################
empirical_ES(model_anchor, DIF=FALSE) # test level stats
empirical_ES(model_anchor) # item level stats
empirical_ES(model_anchor, DIF=FALSE, plot=TRUE) # expected test score plots
empirical_ES(model_anchor, plot=TRUE) # expected item score plots

And the final output (test level):


> empirical_ES(model_anchor, DIF=FALSE) # test level stats
Effect Size Value
1 STDS 1.114
2 UTDS 1.787
3 UETSDS 1.114
4 ETSSD 0.189
5 Starks.DTFR 1.096
6 UDTFR 1.775
7 UETSDN 1.096
8 theta.of.max.test.D 0.513
9 Test.Dmax 1.241

Item level output:


> empirical_ES(model_anchor) # item level stats
SIDS UIDS SIDN UIDN ESSD theta.of.max.D max.D mean.ES.foc mean.ES.ref
item.1 -0.112 0.112 -0.113 0.113 -0.333 2.538 -0.226 0.839 0.952
item.2 0.000 0.000 0.000 0.000 0.000 -0.636 0.000 0.816 0.816
item.3 -0.011 0.030 -0.013 0.032 -0.021 2.265 -0.112 0.750 0.760
item.4 0.013 0.028 0.012 0.030 0.137 2.932 -0.111 0.619 0.606
item.5 -0.014 0.059 -0.013 0.062 -0.027 -2.094 0.138 1.890 1.904
item.6 -0.052 0.059 -0.052 0.061 -0.092 1.790 -0.147 1.000 1.051
item.7 -0.051 0.051 -0.051 0.051 -0.138 0.398 -0.057 0.755 0.806
item.8 0.000 0.000 0.000 0.000 0.000 -0.636 0.000 0.766 0.766
item.9 0.000 0.000 0.000 0.000 0.000 -0.636 0.000 1.135 1.135
item.10 0.000 0.000 0.000 0.000 0.000 -0.636 0.000 0.717 0.717
item.11 -0.053 0.053 -0.051 0.051 -0.211 0.201 -0.060 0.456 0.508
item.12 0.000 0.000 0.000 0.000 0.000 -0.636 0.000 1.167 1.167
item.13 0.466 0.466 0.464 0.464 1.288 2.111 0.698 1.097 0.631
item.14 0.508 0.508 0.498 0.498 1.015 0.391 0.574 1.502 0.994
item.15 0.420 0.420 0.414 0.414 0.808 1.209 0.596 1.127 0.707
>
10

And the ETS plot. Item plots are also output but not presented here.

You might also like