ThurMod
are performed. For the
moment, we separate two different model types:
The Thurstonian factor model was introduced by Maydeu-Olivares & Böckenholt (2005), the Thurstonian IRT model was introduced by Maydeu-Olivares & Brown (2010). For a review see Jansen & Schulze (2023a). For further extensions and discussion about the model types see Jansen & Schulze (2023b). For example, the factor and IRT model both can be further differentiated. For the differentiation we use the design matrix \(A\) which represents all paired comparisons of a design. For a design of four items, this would be
[,1] [,2] [,3] [,4]
[1,] 1 -1 0 0
[2,] 1 0 -1 0
[3,] 1 0 0 -1
[4,] 0 1 -1 0
[5,] 0 1 0 -1
[6,] 0 0 1 -1
The design matrix can be (Jansen & Schulze, 2023b):
First, we will load the package required in the vignette.
The next step is to define the model we are interested in. For this we have to define the following aspects:Further we simulate data of 1000 respondents with loadings between .30 and .95 and latent utility means between -1 and 1. We assume that the data results from rankings, that is that transitivity between responses holds (the variance of the binary indicators is zero). Further, we assume uncorrelated data. We will simulate the data of all paired comparisons possible (see Jansen & Schulze, 2023b).
Set up the simulation conditions:
nfactor <- 4
nitem <- 12
nperson <- 1000
itf <- rep(1:4, 3)
varcov <- diag(1, 4)
# latent utility means
set.seed(69)
lmu <- runif(nitem, -1, 1)
loadings <- runif(nitem, 0.30, 0.95)
Next, we simulate a data set based on the true parameter values:
data <- sim.data(nfactor = nfactor, nitem = nitem, nperson = nperson, itf = itf,
varcov = varcov, lmu = lmu, loadings = loadings)
#save the file
write.table(data, paste0(tempdir(),'/','myDataFile_f.dat'), quote = FALSE,
sep = " ", col.names = FALSE, row.names = FALSE)
The data set contains all (12 )/2 = 66 possible paired comparison variables. With this data set we will perform analyses of the full design (IRT, CFA).
Full designs include all paired comparisons. The estimation of these designs can take a while, as with categorical data a large correlations matrix must be estimated.
For all functions, the blocks we use must be defined. In a full design, there is only one block with all items:
The blocks
must be defined as a matrix where the rows
correspond to each block. Only for a full design, a vector, for example,
1:12
would work.
We will analyse the data with Mplus and lavaan
. We can
do this in two ways: First, in three separate steps, second,
directly.
Way 1, step 1: Build syntax
# Mplus
syntax.mplus(blocks, itf, model = 'lmean', input_path = 'myFC_model_f.inp', data_path =
"myDataFile_f.dat")
For step 2, now these syntaxes must be run (evaluation will not be performed here):
#lavaan
results_lav1 <- lavaan::lavaan(modelsyn, data = data, ordered = TRUE, auto.fix.first = FALSE,
auto.var = TRUE, int.lv.free = TRUE, parameterization = "theta",
estimator = 'ULSMV')
If you replicate this, be patient, it can take about 20 minutes each, depending on your processing power.
Step 3: Read results. For Mplus you can either open the output file,
or use read.mplus
# Mplus
results_mplus1 <- read.mplus(blocks, itf, model = 'lmean', output_path = "myFC_model_f.out")
srmr cfi tli chisq df pvalue
0.0460 1.0000 1.0000 2131.5890 2171.0000 0.7231
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
20861.0240 2145.0000 0.0000 0.0000 1.0000 0.0000
rmsea.ci.upper
0.0060
results_lav1 <- lavaan::fitmeasures(results_lav1)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
2.130962e+03 2.171000e+03 7.261660e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 5.706839e-03 1.000000e+00 1.000000e+00
The second way is to use fit_mplus
or
fit_lavaan
, each function does the above steps at once:
For IRT models, the procedure is the same, just change
model = 'lmean'
to model = 'irt'
, for
example
# Mplus
results_mplus2irt <- fit.mplus(blocks, itf, model = 'irt', input_path = 'myFC_model_irt.inp',
output_path = "myFC_model_irt.out", data_path = "myDataFile_f.dat")
srmr cfi tli chisq df pvalue
0.0460 1.0000 1.0000 2076.5110 2116.0000 0.7261
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
20861.0240 2145.0000 0.0000 0.0000 1.0000 0.0000
rmsea.ci.upper
0.0060
results_lav2irt <- lavaan::fitmeasures(results_lav2irt)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
2.075897e+03 2.116000e+03 7.291136e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 5.719821e-03 1.000000e+00 1.000000e+00
The first block designs were introduced as multidimensional forced-choice blocks (Brown & Maydeu-Olivares, 2011). However, it was shown that the designs must neither be multidimensional, nor must every paired comparison only be contained once (Jansen & Schulze, 2023b).
We first simulate a new data set. Set up the simulation conditions:
nfactor <- 5
nitem <- 30
nperson <- 1000
itf <- rep(1:5, 6)
varcov <- diag(1, 5)
# latent utility means
set.seed(69)
lmu <- runif(nitem, -1, 1)
loadings <- runif(nitem, 0.30, 0.95)
Next, we simulate a data set based on the true parameter values:
set.seed(1234)
data <- sim.data(nfactor = nfactor, nitem = nitem, nperson = nperson, itf = itf, varcov = varcov,
lmu = lmu, loadings = loadings)
#save the file
write.table(data,'myDataFile.dat', quote = FALSE, sep = " ", col.names = FALSE, row.names = FALSE)
Next, we consider unlinked blocks of three items (triplets):
The blocks are defined by a matrix where the rows are the blocks and the number of columns is the number of items per block. Before we can fit the model, we must ensure that the data fits to the syntax we create. The data simulated before assumes that all items are in ascending order. This is important, as the order of the items determine the coding. We code binary items as \[\begin{equation} y_{l} = \begin{cases} 1 & \text{if item $i$ is chosen over item $j$} \\ 0 & \text{else} .\\ \end{cases} \end{equation}\] In a ranking task, all choice alternatives are presented at once, but for each possible comparison, the coding scheme can be used. For example, consider \(n = 3\) items which are labeled as {A, B, C}. Then we have the pairs {A, B}, {A, C}, {B, C}. If for {A, B} A is preferred over B then \(y_{A,B}=1\) and 0 otherwise. This can be done for every paired comparison and ranking task (Maydeu-Olivares & Böckenholt, 2005): For {A,C,B} the binary outcomes are \(y_{A,B}=1\), \(y_{A,C}=1\), and \(y_{B,C}=0\). However, if the order of the binary item is changed, then the coding is changed accordingly, e.g. \(y_{B,A}=0\), \(y_{C,A}=0\), and \(y_{C,B}=1\).
We can get an overview over all possible comparisons by the block
design, by using pair.combn
:
[,1] [,2]
[1,] 1 15
[2,] 2 19
[3,] 4 18
[4,] 4 21
[5,] 5 20
[6,] 7 30
[7,] 8 2
[8,] 8 19
[9,] 9 10
[10,] 11 6
[11,] 13 17
[12,] 14 1
[13,] 14 15
[14,] 16 5
[15,] 16 20
[16,] 21 18
[17,] 22 3
[18,] 23 3
[19,] 23 22
[20,] 24 12
[21,] 24 26
[22,] 25 13
[23,] 25 17
[24,] 26 12
[25,] 27 7
[26,] 27 30
[27,] 28 6
[28,] 28 11
[29,] 29 9
[30,] 29 10
See that we have some items, that are not in ascending order. The
data simulated assumes, that the first named item (item \(i\)) has a smaller number e.g., i3i9. The
blocks we created however, have some comparisons, where the first items
have a larger number e.g. i9i3. There are two ways to fit the syntax to
the data: First, we can simply sort the blocks via
blocksort
:
[,1] [,2]
[1,] 1 14
[2,] 1 15
[3,] 2 8
[4,] 2 19
[5,] 3 22
[6,] 3 23
[7,] 4 18
[8,] 4 21
[9,] 5 16
[10,] 5 20
[11,] 6 11
[12,] 6 28
[13,] 7 27
[14,] 7 30
[15,] 8 19
[16,] 9 10
[17,] 9 29
[18,] 10 29
[19,] 11 28
[20,] 12 24
[21,] 12 26
[22,] 13 17
[23,] 13 25
[24,] 14 15
[25,] 16 20
[26,] 17 25
[27,] 18 21
[28,] 22 23
[29,] 24 26
[30,] 27 30
Now all paired comparisons that can be derived are in ascending order.
The second way is to recode the corresponding binary indicators in the data, as if we flip the coding schema. We can just recode the variables:
# Get names of binary indicators that have non-ascending names
tmp <- which(pair.combn(blocks)[,1] > pair.combn(blocks)[,2])
# get names
pair_names_b <- i.name(blocks)
pair_names_ori <- i.name(1:nitem)
# Rename
pair_names <- i.name(1:nitem)
if(length(tmp) != 0){
tmp1 <- pair_names_b[tmp]
tmp2 <- sub('^i.+i','i', tmp1)
tmp3 <- tmp1
for(j in 1:length(tmp)){
tmp3 <- paste0(tmp2[j], sub(paste0(tmp2[j],'$'), '', tmp1[j]))
pair_names[which(pair_names %in% tmp3)] <- pair_names_b[tmp[j]]
}
}
tmp <- which(!names(data) %in% pair_names)
# Clone data
data_recoded <- data
# Recode and rename
data_recoded[,tmp] <- abs(data[,tmp]-1)
names(data_recoded) <- pair_names
# Save data
write.table(data_recoded, 'myDataFile_rec.dat', quote = FALSE, sep = " ", col.names = FALSE,
row.names = FALSE)
Both ways yield the same results. We have to add the argument
data_full = TRUE
, as we simulated all data in the data
file, but only use some of the data. We demonstrate only the
model = 'irt'
case, however, also for the CFA model types,
these analyses can be performed (model = 'lmean'
,
model = 'uc'
, or model = 'simple'
).
# Blocks_sorted
# Mplus
results_mplus_b <- fit.mplus(blocks_sorted, itf, model = 'irt', input_path = 'myFC_model_bu.inp',
output_path = "myFC_model_bu.out", data_path = "myDataFile.dat",
data_full = TRUE)
srmr cfi tli chisq df pvalue
0.041 1.000 1.000 356.264 375.000 0.749
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
4171.100 435.000 0.000 0.000 1.000 0.000
rmsea.ci.upper
0.009
#lavaan
results_lav_b <- fit.lavaan(blocks_sorted, itf, model = 'irt', data = data)
results_lav_b_fm <- lavaan::fitmeasures(results_lav_b)
results_lav_b <- lavaan::fitmeasures(results_lav_b)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
3.560344e+02 3.750000e+02 7.517818e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 8.660242e-03 1.000000e+00 1.000000e+00
# Recoded data
# Mplus
results_mplus_brec <- fit.mplus(blocks, itf, model = 'irt', input_path = 'myFC_model_brecu.inp',
output_path = "myFC_model_brecu.out", data_path =
"myDataFile_rec.dat", data_full = TRUE)
srmr cfi tli chisq df pvalue
0.041 1.000 1.000 356.264 375.000 0.749
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
4171.100 435.000 0.000 0.000 1.000 0.000
rmsea.ci.upper
0.009
results_lav_brec <- lavaan::fitmeasures(results_lav_brec)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
3.560344e+02 3.750000e+02 7.517822e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 8.660236e-03 1.000000e+00 1.000000e+00
In the case of blocks, we must correct the fit indices, as there are
redundancies among the thresholds and tetrachoric correlations. The
redundancies can be determined with the function
redundancies
. We can also directly correct the fit with
fit.correct
:
#save fit measures
tmp <- results_lav_b_fm
fit.correct(1000, blocks, tmp['chisq.scaled'], tmp['df.scaled'], tmp['baseline.chisq.scaled'],
tmp['baseline.df.scaled'])
df.df_corr rmsea cfi
365 0 1
Now, we consider linked block designs (Jansen & Schulze, 2023b). The simplest way to construct these designs is to take the original unlinked design and link all blocks together (rank of the design \(r_D=N-1\), where \(N\) is the number of items; Jansen & Schulze, 2023b). The original blocks are:
[,1] [,2] [,3]
[1,] 23 22 3
[2,] 14 1 15
[3,] 8 2 19
[4,] 24 26 12
[5,] 28 11 6
[6,] 4 21 18
[7,] 25 13 17
[8,] 16 5 20
[9,] 29 9 10
[10,] 27 7 30
To get a linked design we need extra blocks. The number can be
determined with count.xblocks
[1] 5
Hence, we need five extra triplets in this case. We add for example the
following linking blocks:
blocks_extra <- matrix(c(23,14,8,24,28,4,25,16,29,22,26,13,1,21,30), ncol = 3, byrow = TRUE)
blocks_con <- rbind(blocks, blocks_extra)
We can determine the rank of the design matrix with
rankA
:
[1] 29
The rank is 30-1=29. The function metablock
also gives a
overview over which blocks are linked.
[[1]]
[1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
[29] 29 30
There is only one meta block, therefore all items are linked. A counter example would be if we would not add block 4:
blocks_extra <- matrix(c(23,14,8,24,28,4,25,16,29,1,21,30), ncol = 3, byrow = TRUE)
blocks_con <- rbind(blocks, blocks_extra)
Then the rank is
[1] 28
The rank is 30-1=29\(\neq\) 28. And we have two “meta” blocks.
[[1]]
[1] 5 9 10 13 16 17 20 25 29
[[2]]
[1] 1 2 3 4 6 7 8 11 12 14 15 18 19 21 22 23 24 26 27 28 30
Instead of manually constructing the blocks, ThurMod
enables the user to get extra blocks via a function
get.xblocks
. We have to define if the blocks should be
multidimensional (else multidim = FALSE
), and if items
should not be considered (e.g. because they are negatively keyed):
blocks_extra <- get.xblocks(blocks, itf, multidim = TRUE, item_not = NULL)
blocks_con <- rbind(blocks, blocks_extra)
blocks_con
[,1] [,2] [,3]
[1,] 23 22 3
[2,] 14 1 15
[3,] 8 2 19
[4,] 24 26 12
[5,] 28 11 6
[6,] 4 21 18
[7,] 25 13 17
[8,] 16 5 20
[9,] 29 9 10
[10,] 27 7 30
[11,] 23 14 2
[12,] 12 11 18
[13,] 17 16 9
[14,] 7 23 26
[15,] 20 2 9
If we take the blocks as is, we need to recode the data set:
# Get names of binary indicators that have non-ascending names
tmp <- which(pair.combn(blocks_con)[,1] > pair.combn(blocks_con)[,2])
# get names
pair_names_b <- i.name(blocks_con)
pair_names_ori <- i.name(1:nitem)
# Rename
pair_names <- i.name(1:nitem)
if(length(tmp)!=0){
tmp1 <- pair_names_b[tmp]
tmp2 <- sub('^i.+i','i', tmp1)
tmp3 <- tmp1
for(j in 1:length(tmp)){
tmp3 <- paste0(tmp2[j], sub(paste0(tmp2[j],'$'), '', tmp1[j]))
pair_names[which(pair_names %in% tmp3)] <- pair_names_b[tmp[j]]
}
}
tmp <- which(!names(data) %in% pair_names)
# Clone data
data_recoded <- data
# Recode and rename
data_recoded[,tmp] <- abs(data[,tmp]-1)
names(data_recoded) <- pair_names
# Save data
write.table(data_recoded, 'myDataFile_rec.dat', quote = FALSE, sep = " ", col.names = FALSE,
row.names = FALSE)
The estimation of linked designs is done equivalent via
# Blocks_sorted
# Mplus
results_mplus_bc <- fit.mplus(blocks_con_sorted,itf,model='irt',input_path='myFC_model_b_con.inp',
output_path="myFC_model_b_con.out",data_path="myDataFile.dat",
data_full = TRUE)
srmr cfi tli chisq df pvalue
0.0430 1.0000 1.0000 886.7800 921.0000 0.7858
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
10190.6740 990.0000 0.0000 0.0000 1.0000 0.0000
rmsea.ci.upper
0.0060
#lavaan
results_lav_bc <- fit.lavaan(blocks_con_sorted, itf, model = 'irt', data = data)
results_lav_bc_fm <- lavaan::fitmeasures(results_lav_bc)
results_lav_bc <- lavaan::fitmeasures(results_lav_bc)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
8.863299e+02 9.210000e+02 7.888948e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 6.388203e-03 1.000000e+00 1.000000e+00
# Recoded data
# Mplus
results_mplus_bcrec <- fit.mplus(blocks_con, itf, model = 'irt', input_path = 'myFC_model_brec.inp',
output_path = "myFC_model_brec.out",data_path =
"myDataFile_rec.dat",data_full = TRUE, byblock = FALSE)
srmr cfi tli chisq df pvalue
0.0430 1.0000 1.0000 886.1220 920.0000 0.7835
blchisq bldf blpvalue rmsea rmsea_p rmsea.ci.lower
10190.6740 990.0000 0.0000 0.0000 1.0000 0.0000
rmsea.ci.upper
0.0060
results_lav_bcrec <- lavaan::fitmeasures(results_lav_bcrec)[c('chisq.scaled','df.scaled','pvalue.scaled',
'rmsea.scaled','rmsea.ci.lower.scaled','rmsea.ci.upper.scaled',
'rmsea.pvalue.scaled','cfi.scaled')]
chisq.scaled df.scaled pvalue.scaled rmsea.scaled
8.856718e+02 9.200000e+02 7.866540e-01 0.000000e+00
rmsea.ci.lower.scaled rmsea.ci.upper.scaled rmsea.pvalue.scaled cfi.scaled
0.000000e+00 6.419601e-03 1.000000e+00 1.000000e+00
Here, we also must correct the fit indices, as there are redundancies
among the thresholds and tetrachoric correlations. Again, the
redundancies can be determined with the function
redundancies
. We directly correct the fit with
fit.correct
:
#save fit measures
tmp <- results_lav_bc_fm
fit.correct(1000, blocks_con, tmp['chisq.scaled'], tmp['df.scaled'], tmp['baseline.chisq.scaled'],
tmp['baseline.df.scaled'])
df.df_corr rmsea cfi
906 0 1
The estimation of a full design will most likely seldom be of
interest. More likely, (linked) block designs are of interest.
Especially for simulations with many items (more than 50), many binary
variables are simulated and saved. While the simulation of all items is
important (see Jansen & Schulze, 2023b), saving all items is not.
sim.data
can also return only items that are of relevance.
For that we need to specify the argument variables
. Assume
we have the linked block design we specified before. We only have to
give the item names of the sorted blocks and define it as the
variables
argument:
#Get the relevant names
tmp_names <- i.name(blocks_con_sorted)
#same example as before
set.seed(1234)
data <- sim.data(nfactor = nfactor, nitem = nitem, nperson = nperson, itf = itf, varcov = varcov,
lmu = lmu, loadings = loadings, variables = tmp_names)
#save the file
write.table(data, 'myDataFile.dat', quote = FALSE, sep = " ", col.names = FALSE,
row.names = FALSE)
For the Mplus analysis, the argument data_full
must then
be set to FALSE
.
Taken together, we have demonstrated the basic steps on how to use
the functions given by ThurMod
. Please report any bugs. If
you miss functions that could be helpful for the analysis of Thurstonian
forced-choice data, please let me know.
Brown, A., & Maydeu-Olivares, A. (2011). Item response modeling of forced-choice questionnaires. Educational and Psychological Measurement, 71(3), 460-502. https://doi.org/10.1177/0013164410375112
Jansen, M. T., & Schulze, R. (2023a). Item scaling of social desirability using conjoint measurement: A comparison of ratings, paired comparisons, and rankings. Manuscript in preparation.
Jansen, M. T., & Schulze, R. (in press). The Thurstonian linked bock design: Improving Thurstonian modeling for paired comparison and ranking data. Educational and Psychological Measurement.
Maydeu-Olivares, A., & Böckenholt, U. (2005). Structural equation modeling of paired-comparison and ranking data. Psychological Methods, 10(3), 285–304. https://doi.org/10.1037/1082-989X.10.3.285
Maydeu-Olivares, A., & Brown, A. (2010). Item response modeling of paired comparison and ranking data. Multivariate Behavioural Research, 45(6), 935-974. https://doi.org/10.1080/00273171.2010.531231